In my\u00a0RAG-ING Forward\u00a0sequence\u00a0I \u00a0cloud-native retrieval stack: speech and doc processing, chunking, embeddings, Azure AI Search, and an assistant layer sitting on prime of it. That sequence answered the query I had on the time, which was basically \u2018how do I get a language mannequin to reply questions on paperwork it was by no means skilled on\u2019?\u00a0<\/p>\n
The stack nonetheless works. Retrieval-Augmented Technology stays probably the most sensible solution to floor a mannequin in non-public, domain-specific or just lately modified data with out retraining something.[1] If in case you have a corpus and also you want solutions from it, RAG remains to be the place you begin.\u00a0<\/p>\n
However I’ve been working that sample for some time now, on tasks that lasted longer than a demo, and a unique query began to hassle me:\u00a0<\/p>\n
The system retrieves the identical paragraph, causes over it, produces\u00a0 reply\u00a0\u2014 after which throws\u00a0all\u00a0that reasoning away. Tomorrow, somebody asks a associated query, and it does the equivalent work once more, from scratch, on the similar price, with no assure of reaching the identical conclusion.\u00a0<\/p>\n
My first intuition was to tune the equipment relatively than query it. I experimented with totally different caching mechanisms primarily based on embeddings: recognizing that an incoming query was semantically shut to at least one the system had already answered, and serving the sooner response as an alternative of paying for the total retrieval-and era cross once more. Semantic caching genuinely helps with price and latency, and I’d nonetheless suggest it.\u00a0Nevertheless it took me some time to confess what it really is.\u00a0It\u00a0caches\u00a0solutions, not understanding. The cached response is precisely as disposable as the unique one. Nothing in regards to the\u00a0system\u2019s\u00a0mannequin of the area has improved, and the second a query falls outdoors the similarity threshold, the work begins from zero once more. No matter I tweaked, the primary RAG architectural idea beneath\u00a0remained\u00a0the identical.\u00a0<\/p>\n That isn’t a retrieval downside. Retrieval is doing precisely what it was designed to do. It’s an\u00a0structure\u00a0downside.\u00a0There may be nowhere in a typical RAG system for understanding to build up. No quantity of caching, re-ranking\u00a0or chunking technique fixes that, as a result of all of them\u00a0optimize\u00a0the\u00a0lookup,\u00a0none of them offers the system a reminiscence.\u00a0<\/p>\n This text is about\u00a0constructing\u00a0that\u00a0lacking\u00a0place. It’s the results of my newest work and experimentation round RAG,\u00a0GraphRAG\u00a0and agentic reasoning over a corpus of paperwork, the purpose the place the incremental tweaks stopped being\u00a0sufficient\u00a0and the design itself needed to change. I’ll current it in three elements.\u00a0<\/p>\n Half I is\u00a0vendor-neutral.\u00a0It describes the structure as a design sample: the layers, the thing mannequin, the failure modes it exists to outlive, and the governance it calls for. None of it is determined by Azure, or on any\u00a0explicit database\u00a0or mannequin supplier. If you’re on\u00a0AWS,\u00a0GCP, or working Postgres with\u00a0pgvector\u00a0and a neighborhood mannequin, the design nonetheless holds, and I would love it to be helpful to you.\u00a0<\/p>\n Half II is the Azure implementation.\u00a0Service by service, with the reasoning for every alternative, actual infrastructure-as-code, and a working\u00a0FastAPI\u00a0utility you may clone and deploy.\u00a0<\/p>\n Half III is the demonstration.\u00a0An artificial property insurer known as\u00a0Ostermere\u00a0Mutual, twenty-one interconnected paperwork, and three walkthroughs that present the sample doing one thing a retrieval system genuinely can’t.\u00a0<\/p>\n The whole lot within the dataset is artificial.\u00a0Ostermere\u00a0Mutual doesn’t exist. Neither does the regulator, the coverage, the claims, the folks, the wind\u00a0zones\u00a0or the figures. Nothing right here is insurance coverage, authorized, underwriting or\u00a0claims\u00a0recommendation, and no web page within the demo\u00a0represents\u00a0an actual interpretation of any actual coverage.\u00a0<\/p>\n Naming be aware:\u00a0present Microsoft documentation makes use of\u00a0Microsoft Foundry\u00a0for the unified platform beforehand known as Azure AI Foundry.\u00a0I take advantage of\u00a0Foundry\u00a0all through, whereas protecting the acquainted Azure service names the place they make the structure simpler to comply with.\u00a0<\/p>\n 1. The place basic RAG works, and the place it stops<\/p>\n 2. Retrieval will not be accrued understanding<\/p>\n 3.\u00a0The three\u00a0layers<\/p>\n 4. What\u00a0really lives\u00a0within the information layer<\/p>\n 5.\u00a0The six\u00a0failure modes<\/p>\n 6. Writing information is a unique threat class<\/p>\n 7. Question routing<\/p>\n 8. When you shouldn’t construct this\u00a0<\/p>\n 9. Mapping the layers to providers<\/p>\n 10. Blob Storage<\/p>\n 11. Doc Intelligence<\/p>\n 12. Azure AI Search<\/p>\n 13. Cosmos DB<\/p>\n 14. Microsoft Foundry<\/p>\n 15.\u00a0FastAPI\u00a0on Container Apps<\/p>\n 16. Id<\/p>\n 17. The ingestion lifecycle\u00a0<\/p>\n 18.\u00a0Ostermere\u00a0Mutual<\/p>\n 19. The scoped rule<\/p>\n 20.\u00a0The contradiction<\/p>\n 21.\u00a0The date\u00a0of loss<\/p>\n 22.\u00a0The Obsidian\u00a0vault<\/p>\n 23.\u00a0The fee\u00a0argument<\/p>\n 24. Governance<\/p>\n 25. What I’d construct subsequent<\/p>\n 26. To sum all of it up\u00a0<\/p>\n The whole undertaking (utility, infrastructure,\u00a0dataset\u00a0and a ready-to-open Obsidian vault) is offered within the accompanying GitHub repository at\u00a0github.com\/mcekikj\/persistent-knowledge-layer, below the MIT license.\u00a0<\/p>\n A traditional RAG move is easy sufficient to attract in a single line.\u00a0<\/p>\n Paperwork are chunked, embedded, and saved in a vector-capable index. A query\u00a0arrives,\u00a0the system finds semantically or lexically comparable chunks, and fingers them to the mannequin as context. The mannequin, which is aware of nothing about your corporation, is briefly made to appear to be it does.\u00a0<\/p>\n This solves an actual and essential downside, and I don’t need to\u00a0undersell\u00a0it. The group\u2019s non-public information doesn’t have to dwell within the mannequin\u2019s parameters. It’s fetched when wanted. That could be a genuinely good\u00a0concept\u00a0and it’s why the sample unfold so quick.\u00a0<\/p>\n However have a look at what the structure is\u00a0optimized\u00a0for. It’s\u00a0optimized\u00a0for lookup at question time. Each query is handled as the primary query anybody has ever requested.\u00a0<\/p>\n Now contemplate an actual person, engaged on an actual downside, over a number of weeks:\u00a0<\/p>\n What’s Precise Money Worth?\u00a0<\/p>\n How does it differ from Substitute Price Worth?\u00a0<\/p>\n When can recoverable depreciation be paid?\u00a0<\/p>\n Which earlier choice outlined how we deal with depreciation?\u00a0<\/p>\n Which doc launched the exception, and why?\u00a0<\/p>\n A colleague informed me the edge is 15 years. Is it?\u00a0<\/p>\n A easy RAG utility solutions every of those independently. It retrieves chunks once more, rebuilds context once more, and asks the mannequin to cause once more. Every reply\u00a0might effectively\u00a0be good. However the\u00a0synthesis\u00a0is disposable, that means that when the response is delivered, the understanding evaporates. Nothing in regards to the sixth query is simpler as a result of the system already answered the primary 5.\u00a0<\/p>\n And on the final query \u2013\u00a0is the edge 15 years?\u00a0\u2013 a retrieval system will do one thing worse than\u00a0fail. It would discover the chunk that claims 15 years, and it’ll confidently inform you sure.\u00a0<\/p>\n The analogy I maintain coming again to is a researcher with a submitting cupboard.\u00a0<\/p>\n You ask a query. The researcher goes to the cupboard, pulls 4 paperwork, reads the related passages, and offers you a thought of reply. That is genuinely helpful. Then they put the paperwork again, throw away their notes, and overlook the complete train.\u00a0Tomorrow\u00a0you\u00a0ask\u00a0a follow-up, they usually begin once more on the cupboard.\u00a0<\/p>\n A greater researcher does one thing else whereas studying. They:\u00a0<\/p>\n maintain summaries of the essential sources;\u00a0<\/p>\n keep a web page for every recurring idea, and join the ideas that become associated;\u00a0<\/p>\n write down choices, and the reasoning behind them;\u00a0<\/p>\n replace a comparability when new proof modifications it;\u00a0<\/p>\n document contradictions relatively than quietly resolving them;\u00a0<\/p>\n maintain a working listing of what they nonetheless can’t reply;\u00a0<\/p>\n and maintain a hyperlink from each declare again to the doc it got here from.\u00a0<\/p>\n The primary researcher is a query-time RAG system. The second is what I need to construct.\u00a0<\/p>\n That is near the\u00a0LLM Wiki\u00a0sample that Andrej Karpathy sketched out: uncooked sources keep the place they’re, whereas an agent\u00a0maintains\u00a0a set of Markdown pages (entities, ideas, comparisons, cross-references) {that a} human can learn and navigate. Obsidian occurs to be a handy window onto the consequence.[2]\u00a0<\/p>\n It’s simple to get distracted by the Markdown right here, so let me be exact about what the essential concept\u00a0really is.\u00a0<\/p>\n The essential concept will not be Obsidian. It isn’t even Markdown. It’s that\u00a0information is compiled as soon as right into a sturdy artifact, as an alternative of being reconstructed from uncooked chunks on each request.\u00a0<\/p>\n That’s an architectural declare, and it has architectural penalties.\u00a0<\/p>\n I don’t need to change RAG. I need to give it someplace to place what it learns.\u00a0<\/p>\n The uncooked supply remains to be the strongest factor you will have everytime you want precise wording, a citation, a clause reference, a newly uploaded doc, or verification of one thing disputed. A generated web page,\u00a0nonetheless\u00a0rigorously\u00a0maintained, is a derived interpretation. It isn’t proof, and the second you let it faux to be proof, you will have constructed one thing harmful.\u00a0<\/p>\n So\u00a0the design has three layers, they usually reply three totally different questions.\u00a0<\/p>\n The excellence is sensible, not philosophical. The proof layer is a retrieval index.\u00a0The information layer is a maintained, structured, human-readable mannequin of the area.\u00a0The orchestrator is the factor that is aware of a query about precise coverage wording ought to go to the primary, and a query about\u00a0why we determined this\u00a0ought to go to the second.\u00a0<\/p>\n One rule holds the entire thing collectively, and it’s price\u00a0stating\u00a0by itself line:\u00a0<\/p>\n A information web page isn’t a supply. It’s all the time traceable to at least one.\u00a0<\/p>\n Break that\u00a0rule\u00a0and also you not have a information base. As a substitute,\u00a0you will have a set of assured claims that no one can confirm.\u00a0<\/p>\n The phrase\u00a0wiki\u00a0invitations folks to image a folder of Markdown recordsdata. For a private information base that’s genuinely positive.\u00a0For an utility, I desire a structured retailer beneath, with Markdown as a\u00a0view\u00a0generated from it.\u00a0<\/p>\n The thing mannequin I’ve settled on:\u00a0<\/p>\n Most of those are unsurprising. Three of them are the entire level, and I need to dwell on them.\u00a0<\/p>\n A call object holds the rule, its scope, its efficient date, its accountable proprietor, and its rationale, with a pointer to the place the rationale got here from.\u00a0<\/p>\n That final area issues greater than it appears. In my artificial corpus, an underwriting replace says a roof inspection is triggered above 15 years in a single particular wind zone. The replace doesn’t say\u00a0why\u00a015. The reasoning exists in precisely one place: an e mail thread between an analyst and a head of underwriting, which explains that between 15 and 20 years, roof displacement within the severe-wind band runs about 3.4x the usual band, and which explicitly warns that individuals will learn the quantity with out the qualifier and apply it to the complete ebook.\u00a0<\/p>\n An e mail will not be a coverage doc. No retrieval system ranks it extremely. And in eighteen months, when somebody\u00a0asks\u00a0\u201cwhy is it 15?\u201d, that reasoning is gone \u2014 until one thing intentionally preserved it.\u00a0<\/p>\n That is the concept I’d most encourage you to remove, no matter else you employ from this text.\u00a0<\/p>\n Actual corpora contradict themselves. Two groups write two paperwork, each present, neither superseding the opposite, they usually disagree. In any doc set\u00a0maintained\u00a0by multiple crew for greater than a yr, that is the conventional\u00a0state of affairs, not an edge case.\u00a0<\/p>\n A retrieval system handles this catastrophically badly. It retrieves one chunk, or the opposite, or each, after which asks a language mannequin to reconcile them in a single ahead cross, below a system immediate that informed it to be useful. The mannequin will produce a solution\u00a0that\u00a0might be\u00a0fluent,\u00a0assured,\u00a0however\u00a0it’ll have silently\u00a0picked\u00a0a facet.\u00a0<\/p>\n Worse, the heuristic it most naturally reaches for is\u00a0the more moderen doc wins. That sounds smart, and it’s flawed.\u00a0Recency will not be applicability.\u00a0A more moderen doc might have\u00a0narrower\u00a0scope, might deal with a unique product, or might\u00a0have been written by a crew with no authority over the query.\u00a0<\/p>\n So,\u00a0the contradiction will get its personal object, with a standing, each statements verbatim, their efficient dates, an accountable proprietor, and an express area:\u00a0why_not_resolved. The system\u2019s job is to detect the battle and refuse to settle it.\u00a0<\/p>\n The pure companion. Some questions can’t be answered but, actually because they’re blocked\u00a0by\u00a0contradiction. An open query that’s\u00a0seen\u00a0is protected. The identical query, quietly answered flawed, is the factor that leads to a grievance file.\u00a0<\/p>\n Right here is the trustworthy take a look at of any structure: what does it do this the less complicated factor can’t?\u00a0<\/p>\n I constructed the demonstration corpus particularly to reply that. It\u00a0incorporates\u00a0six distinct traps. A pure retrieval system falls into each certainly one of them \u2014 and\u00a0falls\u00a0fluently, producing a solution that reads completely effectively.\u00a0<\/p>\n A basic rule says\u00a0examine\u00a0roofs\u00a0above\u00a020 years. A later replace says\u00a0examine\u00a0above\u00a015 years, however solely in a single wind zone, and just for new enterprise.\u00a0<\/p>\n Retrieval returns the 15-year chunk. The mannequin says the edge is 15 years. It’s now demanding inspections on tens of hundreds of abnormal roofs, and the\u00a0dealer\u00a0complaints are fully justified.\u00a0<\/p>\n The information layer shops a choice with\u00a0scope: \u201cNew enterprise solely.\u00a0Zone H3 solely.\u201d\u00a0The quantity by no means travels with out its qualifier.\u00a0<\/p>\n The Claims Dealing with Handbook says trace-and-access prices are coated\u00a0as customary\u00a0as much as \u20ac5,000\u00a0and handlers might authorize with out referral. The Endorsement Catalogue says hint and entry is an\u00a0optionally available paid endorsement, not payable until it’s on the schedule.\u00a0<\/p>\n Each paperwork are present. Neither supersedes the opposite. Totally different groups wrote them.\u00a0<\/p>\n A RAG system picks one. The information layer raises a contradiction, names an proprietor, and states that no reply is offered.\u00a0<\/p>\n Throughout the corpus, the identical idea seems as\u00a0precise money worth,\u00a0ACV,\u00a0money settlement\u00a0foundation\u00a0and\u00a0depreciated worth. A\u00a0dealer\u00a0e mail within the dataset\u00a0actually lists\u00a0six such phrases and asks whether or not they’re six issues or one.\u00a0<\/p>\n With out entity decision, your wiki grows 4 separate pages that disagree with one another by omission. With it, one web page, 4 aliases, and a question for any of them lands in the precise place.\u00a0<\/p>\n A declare has a\u00a0date of loss\u00a0of 20 February 2026. A rule took impact on 1 March 2026. The rule can’t apply to that declare.\u00a0<\/p>\n This one is my\u00a0favorite, as a result of a retrieval system has no\u00a0defence\u00a0towards it in any respect. Semantic similarity doesn’t encode time. The chunk in regards to the 15-year threshold is\u00a0maximally related\u00a0to a query about roof age on that declare \u2013 and\u00a0fully inapplicable. The system will not be merely\u00a0flawed,\u00a0it’s flawed in probably the most convincing\u00a0attainable means.\u00a0<\/p>\n The repair requires the orchestrator to know that the query is\u00a0about\u00a0a date, and to pick out the documentation in drive on that date, together with protecting a outdated doc that was dwell on the time.\u00a0<\/p>\n Coated above. The reasoning lives in an\u00a0e mail,\u00a0the rule lives in a suggestion, whereas\u00a0the connection between them lives nowhere.\u00a0<\/p>\n \u201cWhy was this declare triaged Stage 1?\u201d\u00a0requires the declare notes, then the triage guideline, then the water-damage idea, then the coverage clause.\u00a04 hops. Prime-k similarity search doesn’t\u00a0traverse,\u00a0it ranks. Typed relationships\u00a0do\u00a0traverse.\u00a0<\/p>\n Put collectively, these six are the argument \u2014 not that the wiki is nicer, however that there’s a class of query retrieval solutions confidently and wrongly, and the information layer catches it.\u00a0<\/p>\n Right here is the factor that took me longest to internalize, and it modified how I take into consideration the entire design.\u00a0<\/p>\n A flawed chat reply impacts one dialog. A flawed\u00a0canonical\u00a0idea web page impacts each reply that’s later constructed on prime of it, for so long as it stays flawed, and no one notices, as a result of it appears like information.\u00a0<\/p>\n The second your system begins\u00a0writing\u00a0persistent information, it has crossed from\u00a0\u201cretrieval utility\u201d\u00a0into\u00a0\u201csystem of document\u201d,\u00a0and it wants the controls that include that.\u00a0<\/p>\n The mannequin by no means writes to the shop. It proposes a patch. The appliance\u00a0validates\u00a0it and, the place the change is consequential, a human approves it.\u00a0<\/p>\n Observe the place\u00a0\u201cresolve a contradiction\u201d\u00a0sits:\u00a0by no means automated. If the system might resolve contradictions by itself authority, the contradiction object can be pointless.\u00a0<\/p>\n An object whose sources can’t be\u00a0recognized\u00a0ought to be\u00a0deleted, not corrected. You can not repair one thing once you\u00a0don\u2019t\u00a0know the place it got here from.\u00a0<\/p>\n A web page might be\u00a0right\u00a0on Monday and flawed on Friday as a result of a supply was outdated beneath it.\u00a0So\u00a0each derived object carries\u00a0last_validated_at, and the supply it was derived from carries\u00a0superseded_by. When a supply is outdated, every thing derived from it’s marked stale and should not be offered as present till it’s re-derived.\u00a0<\/p>\n Not each query wants each layers, and sending every thing to each is the way you construct one thing costly and gradual.\u00a0<\/p>\n Two issues about this diagram are deliberate.\u00a0<\/p>\n First, the temporal verify belongs earlier than retrieval, not after. In case you rank by similarity first and discard inapplicable outcomes afterwards, the inapplicable paperwork have already consumed your top-k. In manufacturing, put the efficient and outdated dates on the index and filter within the question itself, so the candidate set is constrained earlier than rating. The demo takes a shortcut right here that I ought to come clean with: it applies the date filter\u00a0instantly\u00a0after retrieval, which behaves identically at this corpus measurement however would quietly starve\u00a0top-k\u00a0on a big one. The precept stands\u00a0that\u00a0the demo trades it for a less complicated index schema.\u00a0<\/p>\n Second, the contradiction verify is a gate on the way in which out. It doesn’t matter which route the query took. If the subject is contested, the system\u00a0stops. In code, that’s roughly:\u00a0<\/p>\n def\u00a0question(self, query,\u00a0requested_mode,\u00a0top_k,\u00a0as_of=None):\u00a0 And the instruction that goes to the mannequin is unambiguous:\u00a0<\/p>\n UNRESOLVED CONTRADICTIONS. You MUST current each positions with their sources and\u00a0state that the place is unresolved. You MUST NOT select between them, and also you\u00a0MUST NOT favor the more moderen doc – recency will not be applicability.\u00a0<\/p>\n I’d relatively you skip this structure than misapply it, so let me be direct in regards to the instances the place easy RAG is the higher engineering alternative.\u00a0<\/p>\n Stick with plain RAG when:\u00a0<\/p>\n the corpus is small and queried not often, so there’s nothing to amortize;\u00a0<\/p>\n customers overwhelmingly need precise supply lookup, not synthesis;\u00a0<\/p>\n paperwork churn so quick that any derived synthesis is stale earlier than it’s used;\u00a0<\/p>\n there is no such thing as a cross-session information price preserving;\u00a0<\/p>\n it’s a prototype with a brief life;\u00a0<\/p>\n ingestion latency\u00a0has to\u00a0be minimal;\u00a0<\/p>\n your\u00a0group can’t but govern AI-generated persistent information. This one will not be a technical\u00a0constraint\u00a0and it’s the one folks ignore.\u00a0<\/p>\n Construct the information layer when:\u00a0<\/p>\n the identical area is queried repeatedly, by folks whose work continues throughout classes;\u00a0<\/p>\n cross-source synthesis is regular, not distinctive;\u00a0<\/p>\n choices and their rationale should survive employees turnover;\u00a0<\/p>\n exceptions,\u00a0scopes\u00a0and contradictions\u00a0really matter;\u00a0<\/p>\n area specialists have to\u00a0see\u00a0and proper what the system believes;\u00a0<\/p>\n an audit path from reply to supply is a requirement, not a nice-to-have.\u00a0<\/p>\n This can be a workload choice relatively than a matter of\u00a0new good, outdated dangerous, and the trustworthy reply for lots of functions is that you do not want this.\u00a0<\/p>\n The whole lot above is intentionally moveable. Now let me construct it correctly on Azure, the platform I work with each day.\u00a0<\/p>\n The whole lot else on this structure is derived. Chunks might be re-chunked. Embeddings might be re-embedded. All the wiki can, in precept, be recompiled from scratch. The unique paperwork can’t be recovered from something.\u00a0<\/p>\n So\u00a0they get handled accordingly:\u00a0<\/p>\n raw-sources\/\u00a0 In Bicep, the half that issues\u00a0is\u00a0three properties:\u00a0<\/p>\n useful resource\u00a0blobService\u00a0‘Microsoft.Storage\/storageAccounts\/blobServices@2023-05-01’ = {\u00a0 And on the storage account itself, one line that I’d argue for in any manufacturing deployment:\u00a0<\/p>\n allowSharedKeyAccess: false\u00a0\u00a0 \/\/ no connection strings, ever<\/p>\n The wiki-writing course of must not ever be capable of contact the originals. That could be a provenance, audit and deletion-workflow requirement, and it’s a lot simpler to implement with separate containers and slim function assignments than with good intentions.\u00a0<\/p>\n My demo corpus is\u00a0.txt\u00a0and\u00a0.md, so the appliance reads it immediately. Actual insurance coverage paperwork are scans, kinds, tables and signatures, and the format itself is usually load-bearing. A desk of endorsement limits flattened right into a paragraph is worse than ineffective.\u00a0<\/p>\n Azure AI Doc Intelligence offers you prebuilt and customized fashions that return textual content, tables, choice\u00a0marks\u00a0and construction.[3] My rule of thumb:\u00a0<\/p>\n clear textual content and Markdown \u2192 a easy parser, no cost;\u00a0<\/p>\n machine-readable PDFs \u2192 a PDF parser, whether it is genuinely enough;\u00a0<\/p>\n scans, kinds, complicated format, tables \u2192 Doc Intelligence;\u00a0<\/p>\n all the time\u00a0persist\u00a0the extracted JSON subsequent to the\u00a0unique and\u00a0maintain the web page and span offsets so a quotation can level at a location, not only a doc.\u00a0<\/p>\n That final level pays for itself the primary time a compliance reviewer\u00a0asks\u00a0\u201cthe place precisely does it say that?\u201d\u00a0<\/p>\n Every listed document carries each searchable textual content and its vector:\u00a0<\/p>\n {\u00a0 Azure AI Search matches this design effectively for one particular cause: textual content and vector fields coexist in a single index, and a hybrid question runs the full-text and vector queries in parallel, fusing the rankings with Reciprocal Rank Fusion. Semantic rating can then reorder the highest outcomes.[4]\u00a0<\/p>\n That issues greater than it would sound. In insurance coverage, half the queries are\u00a0conceptual\u00a0(\u201cwhat counts as sudden water injury\u201d) and half are\u00a0lexical\u00a0(\u201cwhat does HS-TA-01 cowl\u201d). Vector search is nice at\u00a0first\u00a0and unreliable at\u00a0second: embeddings are recognized to be weak with precise identifiers, clause\u00a0numbers\u00a0and product codes. BM25 handles these effectively however can’t deal with paraphrasing. You need each, and also you need them fused relatively than chosen between.\u00a0<\/p>\n from\u00a0azure.search.paperwork.fashions\u00a0import\u00a0VectorizedQuery\u00a0 Observe the\u00a0filter, as a result of it’s doing safety work, not relevance tuning.\u00a0<\/p>\n \u26a0\ufe0f Vector similarity will not be authorization.\u00a0Nothing about cosine distance respects your permission mannequin. Safety trimming have to be a tough filter on an listed area, utilized at question time, on each single question. It have to be utilized identically to the information layer and to the exported Markdown. A wiki web page that synthesizes three paperwork the person can’t learn remains to be an information leak, only a properly formatted one.\u00a0<\/p>\n For bigger techniques, built-in vectorization can transfer chunking and embedding into the indexer pipeline.[5] I maintain the embedding calls within the utility right here\u00a0purely,\u00a0so the move is seen and explainable in an article.\u00a0<\/p>\n One sizing be aware from\u00a0really deploying\u00a0this: the demo runs on the\u00a0free\u00a0Search tier, and vector plus hybrid search work positive there for a 21-document corpus. What the free tier offers up is the semantic ranker and managed-identity assist on the service\u00a0itself,\u00a0each are conditionals within the Bicep,\u00a0so deal with\u00a0fundamental\u00a0as the ground for manufacturing and\u00a0free\u00a0as a wonderfully\u00a0great way\u00a0to\u00a0validate\u00a0the design for nothing.\u00a0<\/p>\n Cosmos DB for NoSQL holds the structured wiki. The entire design matches in three choices.\u00a0<\/p>\n Partition\u00a0secret is\u00a0\/workspace_id.\u00a0Each wiki object carries a\u00a0kind\u00a0discriminator, so ideas, choices,\u00a0contradictions\u00a0and open questions all dwell in a single container. Meaning a single-partition question can pull a whole\u00a0workspace\u2019s\u00a0information with no cross-partition fan-out, which is precisely the entry sample this technique has.\u00a0<\/p>\n partitionKey: {\u00a0 No graph database \u2014 but.\u00a0Folks attain for Gremlin or Neo4j the second they hear\u00a0\u201crelationships\u201d. I’d\u00a0push it\u00a0again. Specific relationship information in a doc retailer deal with every thing this technique really\u00a0does:\u00a0discover\u00a0neighbours, comply with a typed edge,\u00a0render\u00a0an idea web page with its hyperlinks. That’s one or two hops.\u00a0<\/p>\n A graph retailer earns its place when deep traversal is\u00a0itself\u00a0the workload: multi-hop influence evaluation, centrality,\u00a0path-finding\u00a0throughout a big community. If you’re not doing that, you might be paying for a second database and a second question language to keep away from writing\u00a0WHERE\u00a0c.source_id\u00a0= @id.\u00a0<\/p>\n Serverless, for now.\u00a0Billing follows consumed request items, which fits a demo and a spiky early workload. Transfer to provisioned throughput upon getting measured actual RU consumption, not earlier than, and never as a result of the phrase\u00a0serverless\u00a0is trendy.[6]\u00a0<\/p>\n Information-plane entry makes use of Cosmos\u2019s personal RBAC system (SQL function assignments), which is separate from Azure RBAC and catches folks out:\u00a0<\/p>\n useful resource\u00a0cosmosDataRole\u00a0‘Microsoft.DocumentDB\/databaseAccounts\/sqlRoleAssignments@2024-11-15’ = {\u00a0 Mixed with\u00a0disableLocalAuth: true, there is no such thing as a key to leak.\u00a0<\/p>\n Foundry offers the chat and embedding deployments. The appliance talks to it via the Azure OpenAI\u00a0v1\u00a0interface utilizing the usual OpenAI Python SDK, which implies no\u00a0api-version\u00a0parameter to chase each few months:[7]\u00a0<\/p>\n from\u00a0openai\u00a0import\u00a0OpenAI\u00a0 For this\u00a0demo,\u00a0the\u00a0FastAPI\u00a0app orchestrates explicitly: embed, retrieve, search the wiki, construct context, name the mannequin,\u00a0validate, write. I did that on objective\u00a0since\u00a0each step is seen, and you may put a breakpoint in any of them.\u00a0<\/p>\n The pure evolution is Foundry Agent Service, the place the identical orchestration turns into an agent with a constrained toolset:[8]\u00a0<\/p>\n search_evidence()\u00a0 I’d not hand an\u00a0agent\u00a0unrestricted database entry on day one. Every software enforces its personal validation, authorization,\u00a0logging\u00a0and slim enter schema. Observe that\u00a0propose_wiki_patch\u00a0and\u00a0apply_approved_patch\u00a0are separate instruments: the agent can attain the primary and can’t attain the second and not using a human in between. That separation is the entire governance mannequin, expressed as an API floor.\u00a0<\/p>\n For extraction, use\u00a0structured outputs, which constrain the response to a JSON Schema as an alternative of merely requesting legitimate JSON.[9] The distinction turns into apparent the primary time a manufacturing extraction returns almost-valid JSON.\u00a0<\/p>\n I can report that from expertise now, as a result of deploying this stack produced two area notes price passing on:\u00a0<\/p>\n Truncated JSON, on the very first doc.\u00a0gpt-5-mini is a reasoning mannequin, and its reasoning tokens are spent from the identical\u00a0max_output_tokens\u00a0price range as the reply. With a 5,000-token cap, the extraction JSON arrived lower off mid-string. The repair within the repo: a 16,000-token price range,\u00a0reasoning: {\u201ceffort\u201d: \u201clow\u201d}\u00a0for extraction work, JSON mode (textual content.format:\u00a0json_object) to constrain the decoder, and one retry. A element that price me two failed seeding runs: the primary full seed succeeded with out JSON mode, then two later runs failed on totally different paperwork. Immediate-only JSON doesn’t fail reliably \u2013 it fails intermittently, which is worse. Schema-constrained structured outputs\u00a0stay\u00a0the actual repair; that is the pragmatic one.\u00a0<\/p>\n A deployment naming quirk.\u00a0An embedding deployment named identically to its mannequin (text-embedding-3-small) got here up wholesome in each standing verify after which returned\u00a0unknown_model\u00a0on each single name, on each the v1 and the basic route. An equivalent deployment named\u00a0embed-3-small\u00a0labored on the primary request. Chat deployments don’t present the issue. The Bicep now retains the deployment identify and the\u00a0mannequin\u00a0identify as separate parameters, with this story within the description.\u00a0<\/p>\n The API floor:\u00a0<\/p>\n GET\u00a0 \/well being\u00a0 One implementation element that price me some debugging time and is price passing on:\u00a0\/wiki\/contradictions\u00a0should\u00a0be declared earlier than\u00a0\/wiki\/{item_id}, or\u00a0FastAPI\u00a0matches the trail parameter first and cheerfully appears for a wiki object with the id\u00a0\u201ccontradictions\u201d. Route order is critical.\u00a0<\/p>\n Container Apps is the precise runtime right here as a result of it’s container-based with out asking me to run Kubernetes.[10] The scaling alternative is price a phrase:\u00a0<\/p>\n scale: {\u00a0 minReplicas\u00a0is a parameter (default\u00a01) as a result of the precise reply is determined by what you might be working. In manufacturing, scale-to-zero appears like free cash after which prices you a chilly begin plus a mannequin handshake on the primary request after each scale-in. For a demo that’s exercised from a developer machine, zero is precisely proper \u2014 the validated deployment on this article ran at zero and price nothing whereas idle.\u00a0<\/p>\n The entire deployment runs on a user-assigned managed id with narrowly scoped roles, and each service has native auth disabled. The Container App will get\u00a0AZURE_CLIENT_ID\u00a0in its\u00a0surroundings,\u00a0DefaultAzureCredential\u00a0picks it up, and no secret is ever issued to the appliance.\u00a0<\/p>\n { identify: ‘AZURE_CLIENT_ID’, worth:\u00a0id.properties.clientId\u00a0}\u00a0<\/p>\n This issues extra on this structure than in a plain RAG app, and it’s price spelling out why. A retrieval system\u00a0reads. This method writes persistent information that different solutions might be constructed on. The blast radius of a compromised credential will not be\u00a0\u201csomebody learn your paperwork\u201d\u00a0\u2013 it’s\u00a0\u201csomebody edited what your group believes\u201d.\u00a0Deal with the write path accordingly.\u00a0<\/p>\n The total\u00a0infra\/fundamental.bicep\u00a0within the repo provisions every thing: storage with versioning, AI Search with\u00a0disableLocalAuth, Cosmos with serverless and SQL function assignments, Foundry with each mannequin deployments, Key Vault, Log Analytics, Software Insights, the Container Apps surroundings and the app itself, plus the six function assignments that wire them collectively \u2013 and, once you cross your personal\u00a0deployerPrincipalId, a mirrored set on your person account, which is what lets the seeding script and the assessments run from a developer machine and not using a single key.\u00a0<\/p>\n az\u00a0group create\u00a0-n\u00a0rg-wikirag-demo\u00a0-l\u00a0swedencentral\u00a0 That element within the center is the one folks miss. When the mannequin extracts ideas from a brand new doc, it should see what the wiki already is aware of.\u00a0In any other case,\u00a0it invents\u00a0\u201cMoney Settlement Foundation\u201d\u00a0as a brand-new idea, with out realizing that\u00a0actual-cash-value\u00a0already exists with that precise alias, and your information base quietly forks.\u00a0<\/p>\n Entity decision within the demo:\u00a0<\/p>\n def\u00a0_resolve_concept(self, identify, aliases):\u00a0 The demo ships with a passing take a look at that ingests a doc utilizing the phrase\u00a0\u201cdepreciated worth\u201d\u00a0and asserts that no new idea web page is created, but it surely resolves onto the present\u00a0actual-cash-value.\u00a0<\/p>\n And here’s what occurs with out the stronger decision steps, measured relatively than argued. Once I ran all twenty-one paperwork via dwell gpt-5-mini extraction on the deployed stack, the mannequin proposed ideas that alias matching might resolve solely partially. The consequence was\u00a0149 idea objects the place the curated graph has 19\u00a0\u2013 a\u00a0roughly 7x\u00a0fragmentation issue. The machine-extracted ideas will not be\u00a0flawed,\u00a0they’re simply named in methods no alias listing\u00a0anticipated\u00a0(\u201cRoof Inspection Requirement\u201d,\u00a0\u201cEUR 5,000 authorization restrict\u201d). That quantity is the concrete argument for embedding similarity and LLM adjudication within the decision chain.\u00a0<\/p>\n The artificial corpus is twenty-one paperwork for an imaginary property insurer. It’s sufficiently small to learn in a day and intentionally engineered in order that the six failure modes from\u00a0stage\u00a05 are all dwell in it.\u00a0<\/p>\n These casual sources will not be\u00a0ornament. The assembly minutes are the place the contradiction is formally recorded as unresolved. The e-mail thread is the place the\u00a0solely\u00a0rationalization of the 15-year threshold exists. In my\u00a0expertise,\u00a0that is precisely how actual organizations\u00a0work;\u00a0the principles are within the paperwork, and the explanations are in somebody\u2019s inbox.\u00a0<\/p>\n A be aware on information provenance: each doc within the corpus is artificial and written by me\u00a0(polished with AI)\u00a0for this text, so there aren’t any licensing constraints on its use. The dataset ships with the repository below the identical MIT license because the code.\u00a0<\/p>\n Compiled, that corpus produces:\u00a0<\/p>\n 21 sources \u00b7 19 ideas \u00b7 28 typed relationships \u00b7 4 comparisons \u00b7 5 choices \u00b7 2 contradictions \u00b7 4 open questions \u00b7 2 processes\u00a0<\/p>\n curl\u00a0-X\u00a0POST http:\/\/localhost:8000\/question\u00a0-H\u00a0‘Content material-Sort: utility\/json’\u00a0-d\u00a0‘{\u00a0 Plain retrieval finds the H3 replace, which says 15 years, and says\u00a0\u201c15 years\u201d.\u00a0<\/p>\n The information layer returns the choice object:\u00a0<\/p>\n {\u00a0 The quantity by no means seems with out its scope. And the rationale, recovered from an e mail, is correct there, which signifies that in eighteen months, when somebody asks\u00a0why 15, the reply exists.\u00a0<\/p>\n The e-mail\u2019s writer, by the way, predicted this precise failure in writing:\u00a0<\/p>\n \u201cPlease don’t let this turn out to be a basic 15-year rule. If somebody reads the replace with out the zone qualifier they’ll apply it to the entire ebook, we are going to demand inspections on tens of hundreds of completely abnormal roofs, and the dealer complaints might be fully justified.\u201d\u00a0<\/p>\n That could be a artificial quote I wrote to make some extent, however I don’t assume it’s an unrealistic one.\u00a0<\/p>\n That is the one I’d put in entrance of a\u00a0sceptical\u00a0architect.\u00a0<\/p>\n curl\u00a0-X\u00a0POST http:\/\/localhost:8000\/question\u00a0-H\u00a0‘Content material-Sort: utility\/json’\u00a0-d\u00a0‘{\u00a0 A retrieval system solutions this. It retrieves a bit (from the claims guide, or from the endorsement catalogue) and tells you both\u00a0\u201csure, as much as \u20ac5,000 as customary\u201d\u00a0or\u00a0\u201cprovided that you purchased HS-TA-01\u201d. Each solutions are supported by an actual doc. Each are\u00a0flawed, as a result of\u00a0the agency doesn’t have a place.\u00a0<\/p>\n The hybrid system responds:\u00a0<\/p>\n {\u00a0 Each statements. Each sources. Each dates. An accountable human being. And no\u00a0reply, as a result of\u00a0there’s\u00a0no\u00a0one.\u00a0<\/p>\n That JSON\u00a0is demo mode\u2019s deterministic scaffold. Here’s what the deployed stack does with the identical query \u2013 gpt-5-mini, dwell, over Cosmos DB and AI Search:\u00a0<\/p>\n There may be an UNRESOLVED CONTRADICTION related to this query. Don’t deal with the difficulty as settled.\u00a0<\/p>\n INS-SYN-012 (Chapter 7.3, efficient 2025-11-01):\u00a0\u201cHint-and-access prices are coated as customary below\u00a0Hearthmere, as much as EUR 5,000. Handlers ought to authorize affordable expenditure as much as this restrict with out referral.\u201d\u00a0<\/p>\n INS-SYN-008 (HS-TA-01, efficient 2026-01-01):\u00a0\u201cHint and entry is an optionally available endorsement (HS-TA-01), restrict EUR 5,000, indicative premium EUR 24. The place it doesn’t seem on the coverage schedule, trace-and-access prices will not be payable.\u201d\u00a0<\/p>\n This contradiction is UNRESOLVED. Accountable proprietor: Y. Tanaka (Product) [con-001].\u00a0<\/p>\n I can’t resolve or state whether or not trace-and-access is roofed below\u00a0Hearthmere\u00a0due to the unresolved battle above.\u00a0<\/p>\n The mannequin was not requested to be cautious typically. It was\u00a0handed\u00a0the contradiction object and one rule, and it adopted the rule.\u00a0<\/p>\n The second-order impact is the half I discover genuinely fascinating. As soon as a contradiction is a saved object, different issues can\u00a0rely\u00a0on it. The corpus\u00a0incorporates\u00a0a policyholder query on declare CLM-1155 (is the leak-detection survey price payable?)\u00a0which\u00a0can’t be answered till this battle is\u00a0resolved.\u00a0So,\u00a0it’s saved as an open query with\u00a0blocked_by: \u201ccon-001\u201d.\u00a0<\/p>\n The system is aware of why it can’t reply, and it is aware of who\u00a0has to\u00a0determine earlier than it will probably.\u00a0<\/p>\n Declare\u00a0CLM-1108: storm injury, roof overlaying displaced,\u00a0date of loss 20 February 2026. The roof is eighteen years outdated. The property, below at the moment\u2019s classification, can be in zone H3.\u00a0<\/p>\n Ask a retrieval system whether or not an inspection was\u00a0required, and it’ll discover the H3 rule (15 years, and 18 > 15) and inform you the roof was over\u00a0threshold,\u00a0and no inspection was on file. It seems like a discovering. It’s the starting of a wrongful declinature.\u00a0<\/p>\n As a result of the H3 rule took impact on\u00a01 March 2026. 9 days\u00a0after\u00a0the loss. And the coverage was written in September 2025, earlier than the zone mannequin existed in any respect, so the property carried no zone classification on the day the roof got here off.\u00a0<\/p>\n curl\u00a0-X\u00a0POST http:\/\/localhost:8000\/question\u00a0-H\u00a0‘Content material-Sort: utility\/json’\u00a0-d\u00a0‘{\u00a0 The\u00a0as_of\u00a0parameter is the date the query is\u00a0about\u00a0\u2013 not the date it’s requested. With it set,\u00a0INS-SYN-004\u00a0is excluded from the candidate set fully. It was not in drive. The relevant threshold was the overall 20-year\u00a0rule,\u00a0the roof was 18 years outdated, it was\u00a0below\u00a0threshold, and no requirement was breached.\u00a0<\/p>\n The identical mechanism runs within the different course. Ask about settlement foundation with\u00a0as_of: \u201c2025-08-01\u201d\u00a0and the system retrieves the outdated coverage v1.1, as a result of that’s the model that ruled a loss on that date, and marks it:\u00a0<\/p>\n {\u00a0 Below v1.1, depreciation was\u00a0not\u00a0recoverable. Below at the moment\u2019s v1.2, it’s. Similar time period, similar coverage, reverse reply \u2013 and the one factor that distinguishes them is a date that similarity search can’t see.\u00a0<\/p>\n Each\u00a0behaviours\u00a0are coated by assessments within the repo, and each\u00a0reproduce on the\u00a0dwell deployment. At\u00a0as_of: 2026-02-20\u00a0the deployed mannequin answered:\u00a0<\/p>\n Quick reply: The final >20-year inspection threshold utilized. [INS-SYN-003]\u00a0<\/p>\n Why: On 2026-02-20\u00a0the\u00a0Hearthmere\u00a0guideline required an inspection when the first roof overlaying is greater than 20 years outdated [INS-SYN-003]. The H3 15-year threshold \u2026 didn’t take impact till 1 March 2026; this coverage was\u00a0written\u00a014 Sep 2025 (so the H3 rule didn’t apply to it) and the roof was ~18 years outdated at loss. [INS-SYN-018] [INS-SYN-013]\u00a0<\/p>\n And at\u00a0as_of: 2025-08-01:\u00a0<\/p>\n Settlement foundation: Precise Money Worth (ACV) \u2013 outlined as alternative price much less depreciation \u2013 for eligible constructing claims below\u00a0Ostermere\u00a0Hearthmere\u00a0v1.1. [INS-SYN-009]\u00a0<\/p>\n Is depreciation recoverable? No.\u202fBelow v1.1 depreciation will not be recoverable after restore. [INS-SYN-009]\u00a0<\/p>\n They’re the 2 assessments I’d maintain if I needed to\u00a0delete\u00a0each different one.\u00a0<\/p>\n Part 5.6 claimed that typed relationships let the system traverse the place similarity search solely ranks. I had not scripted a demo for it, so I requested the deployed stack immediately:\u00a0\u201cClarify why the CLM-1042 water loss was positioned at triage Stage 1 below the property declare triage course of.\u201d\u00a0The reply walked the entire chain:\u00a0<\/p>\n Governing\u00a0Stage-1\u00a0standards: Stage 1 \u2013 Normal is outlined as\u00a0\u201cMinor seen injury, no security concern, no persevering with water entry, and estimated publicity under EUR 5,000\u201d\u00a0[INS-SYN-002].\u00a0<\/p>\n How CLM-1042 matches these standards: property remained liveable, no reported electrical fault [INS-SYN-005]; water move stopped inside roughly 40 minutes [INS-SYN-005]; injury confined to the water path [INS-SYN-005; INS-SYN-014]; preliminary restore estimate EUR 3,800, below EUR 5,000 [INS-SYN-005; INS-SYN-002].\u00a0<\/p>\n Supporting operational steering: the Claims Dealing with Handbook lists indicators of a sudden escape of water \u2026 which align with CLM-1042\u2019s details [INS-SYN-012]. CLM-1155 explicitly contrasts CLM-1042 as a datable, sudden failure [INS-SYN-014].\u00a0<\/p>\n Observe on provisional standing: protection and exclusions\u00a0stay\u00a0topic to a proper protection dedication by a licensed claims skilled [INS-SYN-005].\u00a0<\/p>\n 4 sources, 4 hops, and the human-decision boundary preserved on the finish with out being requested for.\u00a0<\/p>\n One trustworthy statement from the identical session: my first phrasing of that query (\u201c\u2026and the way does the water injury cowl apply to it?\u201d) bought\u00a0gated by the contradiction\u00a0\u2013 CLM-1042\u2019s proof touches the contested trace-and-access idea, so the system offered con-001 and declined to make a settled dedication as an alternative of answering the triage half. The gate\u00a0errs\u00a0on the facet of stopping. For a claims workflow I contemplate that the precise default, however it’s a actual trade-off: an aggressive gate will typically maintain again a solution the person legitimately\u00a0wanted and\u00a0tuning that boundary is a part of\u00a0working\u00a0the system.\u00a0<\/p>\n The appliance exports the complete Cosmos DB state right into a vault. Fifty-five pages, all generated, nothing hand-written:\u00a0<\/p>\n obsidian_vault\/\u00a0 Open the folder in\u00a0Obsidian,\u00a0and you may navigate hyperlinks, examine backlinks, comply with the graph, spot orphan pages, and, most significantly, see what the system believes and inform it that it’s flawed.\u00a0<\/p>\n That final functionality is, I feel, the strongest argument for this complete structure. A vector index is operationally glorious and fully opaque to a site skilled. You can not hand an underwriter a 1,536-dimension embedding and\u00a0ask,\u00a0\u201cdoes this look proper to you?\u201d\u00a0You\u00a0can\u00a0hand them a Markdown web page that claims\u00a0the edge is 15 years, however solely in H3, and just for new enterprise, and right here is why, and listed here are the 4 paperwork it got here from.\u00a0<\/p>\n They are going to inform you inside thirty seconds whether or not it’s proper.\u00a0Markdown makes the reminiscence auditable by the individuals who really know the area.\u00a0No different a part of the stack does that.\u00a0<\/p>\n The information layer prices\u00a0extra at\u00a0ingestion. I’m not going to faux in any other case.\u00a0<\/p>\n A RAG pipeline extracts and embeds every doc as soon as. The hybrid pipeline moreover summarizes, extracts ideas and claims, resolves entities towards the present wiki, generates relationship and comparability patches,\u00a0validates, and regenerates Markdown. The write amplification is\u00a0actual,\u00a0and it’s not small.\u00a0<\/p>\n The argument is that this front-loaded price\u00a0buys down\u00a0repeated query-time reasoning.\u00a0So,\u00a0the query will not be whether or not the wiki prices extra to construct (it plainly does) however whether or not the compilation is amortized throughout sufficient future use.\u00a0<\/p>\n The mannequin in\u00a0cost_model\/cost_model.py:\u00a0<\/p>\n Compilation\u00a0= D \u00d7 Td \u00d7 M\u00a0 With the illustrative defaults \u2013 50 paperwork, 6,000 tokens every, 1,000 questions, 6,000 retrieved context tokens per RAG query versus 1,350 wiki context tokens, uncooked verification on 25% of questions:\u00a0<\/p>\n Now the\u00a0caveats, as a result of\u00a0a chart like this will simply mislead:\u00a0<\/p>\n That is token quantity, not value. It ignores output tokens, embedding prices, AI Search capability, Cosmos RUs, Container Apps compute, and Doc Intelligence pages.\u00a0<\/p>\n It treats each token as equal. In follow, ingestion and answering can use totally different mannequin lessons, and that’s the place a lot of the actual saving lives, as a result of a small mannequin can reply from concise wiki context whereas a stronger one is reserved for reconciliation and updates.\u00a0<\/p>\n It doesn’t assure something. A badly ruled agent that rewrites the entire wiki on each ingestion will erase any saving you modelled. The economics rely fully on disciplined replace insurance policies.\u00a0<\/p>\n One measured information level, from deploying this precise stack: ingesting all twenty-one paperwork via dwell gpt-5-mini extraction and embeddings, working each walkthrough on this article, and regenerating the vault from Cosmos DB price\u00a0roughly\u00a0$0.20-0.30\u00a0in whole. The idle stack \u2013 free-tier search, scale-to-zero Container App, serverless Cosmos \u2013 burns about $0.05 a day. At this corpus\u00a0measurement\u00a0the compilation price is espresso cash.\u00a0The economics\u00a0solely turn out to be fascinating at scale, which is what the mannequin above is for.\u00a0<\/p>\n The context-size assumption additionally survived contact with the deployed system. Measured throughout a set of conceptual questions on the dwell stack, the wiki context averaged\u00a0roughly 500\u00a0tokens towards\u00a0roughly 1,750\u00a0for the equal proof context \u2013 a\u00a03.5x ratio, in the identical vary because the 4.4x the mannequin assumes. Absolutely the numbers are smaller than the\u00a0mannequin\u2019s, as a result of the artificial paperwork are quick. The ratio is the half that transfers to an actual corpus. If something, the measured ratio is barely extra conservative than the assumed one, which strikes the break-even later, not earlier \u2014 price realizing earlier than you quote the mannequin at a price range assembly.\u00a0<\/p>\n Run\u00a0\/cost-estimate\u00a0with your personal assumptions. Then throw them away and use telemetry out of your precise corpus and question combine, priced with the present Azure calculator.[11]\u00a0<\/p>\n And actually, the token argument is the\u00a0weakest\u00a0argument for this structure. The actual returns are:\u00a0<\/p>\n constant terminology throughout classes and throughout folks;\u00a0<\/p>\n choices and rationale that survive the one who made them\u00a0depart;\u00a0<\/p>\n contradictions which are seen as an alternative of silently resolved;\u00a0<\/p>\n an audit path from any reply to its supply;\u00a0<\/p>\n a information artifact a site skilled can\u00a0overview;\u00a0<\/p>\n continuity throughout agent classes and throughout mannequin upgrades.\u00a0<\/p>\n I’d\u00a0take\u00a0over a couple of million enter tokens.\u00a0<\/p>\n As a result of the system writes, it wants\u00a0to regulate\u00a0a read-only system doesn’t.\u00a0<\/p>\n Protect provenance.\u00a0Each assertion\u00a0traces to\u00a0a bit, to a span, to a doc, on the model it was derived from.\u00a0<\/p>\n Patch, by no means write.\u00a0The mannequin\u00a0proposes;\u00a0deterministic validation and, for something consequential, a human disposes.\u00a0<\/p>\n Detect staleness.\u00a0last_validated_at,\u00a0superseded_by,\u00a0source_effective_date. When a supply is outdated, every thing derived from it’s stale till re-derived.\u00a0<\/p>\n Trim safety at each layer.\u00a0Blob paths, search information, Cosmos objects, Markdown exports, agent instruments, caches, telemetry. Constantly. A synthesized web page must not ever floor data the reader couldn’t entry within the supply.\u00a0<\/p>\n Maintain\u00a0the human choice human.\u00a0Within the demo, the AI might summarize an consumption, retrieve coverage proof,\u00a0determine\u00a0lacking data, suggest a provisional triage\u00a0stage\u00a0and flag conflicting guidelines. It could not\u00a0decide\u00a0protection, decline a declare, assess fraud, charge a threat, resolve a contradiction, or inform a\u00a0buyer\u00a0a choice has been made.\u00a0<\/p>\n The artificial working-group minutes\u00a0comprise\u00a0the most effective articulation of this that I managed to put in writing, and I’ll let it stand because the governance precept for the entire structure:\u00a0<\/p>\n Handlers at present undertake the assistant\u2019s proposed triage stage in\u00a0roughly 90%\u00a0of instances, which is ok, however solely as a result of they’re studying the consumption themselves. Eradicating the handler from the loop removes the factor that makes the 90% reliable.\u00a0<\/p>\n That’s the lure in a single sentence. An automatic system earns credibility below human overview, after which that credibility is used because the argument for eradicating the overview.\u00a0<\/p>\n The repo is a baseline, not a product. The gaps I’m most aware of:\u00a0<\/p>\n Occasion Grid and queue-driven async ingestion (the demo ingests synchronously as a result of it’s simpler to run domestically).\u00a0<\/p>\n Doc Intelligence with web page and span preservation, so citations level at places.\u00a0<\/p>\n Strict JSON Schema structured outputs on each extraction and patch.\u00a0<\/p>\n Entity decision with embedding similarity and LLM adjudication, not simply\u00a0alias\u00a0matching.\u00a0<\/p>\n A human approval UI for high-risk patches: proper\u00a0now\u00a0the lifecycle exists within the\u00a0design\u00a0and the low-risk path exists within the code.\u00a0<\/p>\n Foundry Agent Service instruments, with\u00a0suggest\u00a0and\u00a0apply\u00a0as\u00a0separately-permissioned\u00a0surfaces.\u00a0<\/p>\n Analysis units for retrieval, synthesis and the onerous one: replace accuracy. How do you take a look at {that a} information base modified\u00a0accurately?\u00a0<\/p>\n Freshness and\u00a0contradiction\u00a0dashboards. A contradiction\u00a0register\u00a0no one appears at is only a log file.\u00a0<\/p>\n Per-tenant safety trimming, finish to finish.\u00a0<\/p>\n Mannequin routing by job complexity and threat.\u00a0<\/p>\n Quantity 7 is the genuinely open analysis downside, and I shouldn’t have\u00a0 reply\u00a0to it but.\u00a0<\/p>\n RAG is the proof engine of this structure, and nothing right here is its obituary. It offers the mannequin entry to unique, related, present supply materials, and there’s no substitute for that.\u00a0<\/p>\n What it doesn’t do is keep in mind.\u00a0<\/p>\n The information layer provides the factor that was lacking: a maintained, structured, inspectable illustration of what the system has already labored out, with the scopes intact, the rationale preserved, the contradictions seen, and a line again to the proof for each declare.\u00a0<\/p>\n RAG asks:\u00a0What ought to I retrieve for this query?\u00a0<\/p>\n The information layer: What ought to be durably true after processing every thing thus far?\u00a0<\/p>\n The orchestrator: Which of these do I have to reply this safely, proper now?\u00a0<\/p>\n On Azure that separation maps cleanly:\u00a0<\/p>\n Blob Storage\u00a0preserves the originals \u2014 the one factor you can not regenerate.\u00a0<\/p>\n Doc Intelligence\u00a0extracts the troublesome content material and retains the spans.\u00a0<\/p>\n Azure AI Search\u00a0shops retrievable, security-trimmed proof.\u00a0<\/p>\n Cosmos DB\u00a0shops the evolving ideas, relationships,\u00a0choices\u00a0and contradictions.\u00a0<\/p>\n Microsoft Foundry\u00a0offers the fashions, and the trail to brokers.\u00a0<\/p>\n FastAPI\u00a0on Container Apps\u00a0runs the orchestration, the temporal\u00a0scoping\u00a0and the contradiction gate.\u00a0<\/p>\n Obsidian\u00a0makes the entire thing seen to the individuals who know whether or not\u00a0it\u2019s\u00a0proper.\u00a0<\/p>\n It’s extra work than easy RAG, and it prices extra to ingest. In trade you get organizational reminiscence, constant synthesis, express relationships, a visual choice historical past, and, the half I maintain coming again to, a system that can inform you\u00a0\u201ctwo\u00a0of our paperwork disagree and no one has determined but\u201d\u00a0as an alternative of confidently making one thing up.\u00a0<\/p>\n That final functionality will not be a characteristic. It’s the cause to construct it.\u00a0<\/p>\n The appliance is not\u00a0looking out\u00a0a pile of paperwork. It’s slowly constructing a reviewable mannequin of a site, whereas protecting the unique proof shut sufficient to verify each essential conclusion towards.\u00a0<\/p>\n Thanks for taking the time to discover this structure with me. It was an extended piece than I supposed, as a result of the design stored having another half price explaining. The\u00a0FastAPI\u00a0undertaking, the Bicep templates, the twenty-one artificial paperwork and the Obsidian vault are all in\u00a0the repository, and\u00a0I strongly consider they\u00a0offer you a sensible place to begin for constructing a persistent information layer of your personal. Clone it, run it in demo mode with none Azure credentials, and attempt to break it \u2014 I’d genuinely like to listen to the place it fails.\u00a0<\/p>\n Disclosure: I’m a Microsoft MVP. This text displays my very own unbiased work and opinions; Microsoft\u00a0had no involvement\u00a0in or overview of its content material. All Azure utilization described is predicated on public documentation and my very own deployment.\u00a0<\/p>\n [1] P. Lewis et al.,\u00a0Retrieval-Augmented Technology for Information-Intensive NLP Duties\u00a0(2020),\u00a0NeurIPS\u00a02020\u00a0<\/p>\n [2] A. Karpathy,\u00a0LLM Wiki\u00a0(2025), GitHub Gist\u00a0<\/p>\n [3] Microsoft,\u00a0Doc Intelligence Structure Mannequin\u00a0(2026), Microsoft Be taught\u00a0<\/p>\n [4] Microsoft,\u00a0Hybrid Search Overview \u2013 Azure AI Search\u00a0(2026), Microsoft Be taught\u00a0<\/p>\n [5] Microsoft,\u00a0Built-in Vectorization in Azure AI Search\u00a0(2026), Microsoft Be taught\u00a0<\/p>\n [6] Microsoft,\u00a0Azure Cosmos DB Serverless\u00a0(2026), Microsoft Be taught\u00a0<\/p>\n [7] Microsoft,\u00a0Azure OpenAI v1 API Lifecycle\u00a0(2026), Microsoft Be taught\u00a0<\/p>\n [8] Microsoft,\u00a0Foundry Agent Service Overview\u00a0(2026), Microsoft Be taught\u00a0<\/p>\n [9] Microsoft,\u00a0Structured Outputs with Azure OpenAI\u00a0(2026), Microsoft Be taught\u00a0<\/p>\n [10] Microsoft,\u00a0Deploy a Flask or FastAPI Net App on Azure Container Apps\u00a0(2026), Microsoft Be taught\u00a0<\/p>\n [11] Microsoft,\u00a0Plan and Handle Prices of Azure AI Search\u00a0(2026), Microsoft Be taught\u00a0<\/p>\n<\/div>\n![\"\"](\"https:\/\/contributor.insightmediagroup.io\/wp-content\/uploads\/2026\/07\/image-287-465x1024.png\")
Contents\u00a0<\/h2>\n
Half I \u2013 The design:\u00a0<\/h3>\n
Half II \u2013 Implementing it on Azure:\u00a0<\/h3>\n
Half III \u2013 The demonstration:\u00a0<\/h3>\n
Half I \u2013 The design\u00a0<\/h2>\n
1. The place basic RAG works, and the place it stops\u00a0<\/h2>\n
![\"\"](\"https:\/\/contributor.insightmediagroup.io\/wp-content\/uploads\/2026\/07\/image-290-1024x49.png\")
2. Retrieval will not be accrued understanding\u00a0<\/h2>\n
3. The three layers\u00a0<\/h2>\n
![\"\"](\"https:\/\/contributor.insightmediagroup.io\/wp-content\/uploads\/2026\/07\/image-282-1024x707.png\")
4. What\u00a0really lives\u00a0within the information layer\u00a0<\/h2>\n
![\"\"](\"https:\/\/contributor.insightmediagroup.io\/wp-content\/uploads\/2026\/07\/image-288-1024x993.png\")
Determination \u2013 as a result of the\u00a0why\u00a0dies first\u00a0<\/h3>\n
Contradiction \u2013 a first-class object, not an error\u00a0<\/h3>\n
Open query \u2013 realizing what you\u00a0don\u2019t\u00a0know\u00a0<\/h3>\n
5. The six failure modes this structure exists to outlive\u00a0<\/h2>\n
5.1 Scoped supersession\u00a0<\/h3>\n
5.2 Real contradiction\u00a0<\/h3>\n
5.3 Terminology drift\u00a0<\/h3>\n
5.4 Efficient-date scoping\u00a0<\/h3>\n
5.5 Rationale loss\u00a0<\/h3>\n
5.6\u00a0Multi-hop\u00a0<\/h3>\n
6. Writing information is a unique threat class than answering\u00a0<\/h2>\n
The whole lot is a patch\u00a0<\/h3>\n
![\"\"](\"https:\/\/contributor.insightmediagroup.io\/wp-content\/uploads\/2026\/07\/image-289-780x1024.png\")
Provenance is a sequence, not a area\u00a0<\/h3>\n
![\"\"](\"https:\/\/contributor.insightmediagroup.io\/wp-content\/uploads\/2026\/07\/image-281-1024x78.png\")
Staleness is a property you need to monitor\u00a0<\/h3>\n
7. Routing:\u00a0which\u00a0layer solutions this?\u00a0<\/h2>\n
![\"\"](\"https:\/\/contributor.insightmediagroup.io\/wp-content\/uploads\/2026\/07\/image-291-503x1024.png\")
\n\u00a0\u00a0\u00a0 mode\u00a0=\u00a0self.choose_mode(query,\u00a0requested_mode)\u00a0
\n\u00a0
\n\u00a0\u00a0\u00a0\u00a0wiki_items\u00a0=\u00a0self._search_wiki(query,\u00a0top_k)\u00a0if\u00a0mode\u00a0in\u00a0{“wiki”,\u00a0“hybrid”}\u00a0else\u00a0[]\u00a0
\n\u00a0\u00a0\u00a0 proof\u00a0=\u00a0self.proof.search(query,\u00a0top_k)\u00a0if\u00a0mode\u00a0in\u00a0{“proof”,\u00a0“hybrid”}\u00a0else\u00a0[]\u00a0
\n\u00a0
\n\u00a0\u00a0\u00a0\u00a0if\u00a0as_of:\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0# The date the query is ABOUT – not the date it’s requested.\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 proof\u00a0=\u00a0self._filter_by_date(proof,\u00a0as_of)\u00a0
\n\u00a0
\n\u00a0\u00a0\u00a0\u00a0# Pull in each contradiction touching a retrieved idea, even when the\u00a0
\n\u00a0\u00a0\u00a0\u00a0# contradiction object itself didn’t rank. Somebody asking about hint and\u00a0
\n\u00a0\u00a0\u00a0\u00a0# entry will get the battle whether or not or not they used the phrase “contradiction”.\u00a0
\n\u00a0\u00a0\u00a0 contradictions\u00a0=\u00a0self._contradictions_for(wiki_items)\u00a0
\n\u00a0\u00a0\u00a0 warnings\u00a0=\u00a0self._warnings(proof, contradictions,\u00a0as_of)\u00a0
\n\u00a0
\n\u00a0\u00a0\u00a0 context\u00a0=\u00a0self._build_context(wiki_items, proof, contradictions,\u00a0as_of)\u00a0
\n\u00a0\u00a0\u00a0 reply\u00a0=\u00a0self.mannequin.reply(query, context)\u00a0
\n\u00a0\u00a0\u00a0 …\u00a0<\/p>\n8. When it is best to\u00a0not\u00a0construct this\u00a0<\/h2>\n
Half II \u2013 Implementing it on Azure\u00a0<\/h2>\n
9. Mapping the layers to providers\u00a0<\/h2>\n
![\"\"](\"https:\/\/contributor.insightmediagroup.io\/wp-content\/uploads\/2026\/07\/image-293-1024x810.png\")
![\"\"](\"https:\/\/contributor.insightmediagroup.io\/wp-content\/uploads\/2026\/07\/image-285.png\")
10. Blob Storage \u2013 the one factor you can not regenerate\u00a0<\/h2>\n
\n\u00a0\u00a0\u00a0 {workspace-id}\/\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 {document-id}\/\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 original-file.pdf\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 \u2190 by no means rewritten\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 extracted-content.json\u00a0\u00a0\u00a0\u00a0 \u2190 Doc Intelligence output\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 ingestion-metadata.json\u00a0\u00a0\u00a0 \u2190 what ran, when, which mannequin model\u00a0
\n\u00a0
\nwiki-export\/\u00a0
\n\u00a0\u00a0\u00a0 {workspace-id}\/\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 Residence.md\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 Ideas\/ \u00b7 Choices\/ \u00b7 Contradictions\/ \u00b7 Sources\/\u00a0<\/p>\n
\n\u00a0 mum or dad: storage\u00a0
\n\u00a0 identify: ‘default’\u00a0
\n\u00a0 properties: {\u00a0
\n\u00a0\u00a0\u00a0 \/\/ Originals should survive an ingestion bug. Versioning and mushy delete are the\u00a0
\n\u00a0\u00a0\u00a0 \/\/ most cost-effective insurance coverage obtainable on the one artifact the system can’t regenerate.\u00a0
\n\u00a0\u00a0\u00a0\u00a0isVersioningEnabled: true\u00a0
\n\u00a0\u00a0\u00a0\u00a0deleteRetentionPolicy:\u00a0{ enabled: true, days: 30 }\u00a0
\n\u00a0\u00a0\u00a0\u00a0containerDeleteRetentionPolicy:\u00a0{ enabled: true, days: 30 }\u00a0
\n\u00a0 }\u00a0
\n}\u00a0<\/p>\n11. Doc Intelligence \u2013 used selectively\u00a0<\/h2>\n
12. Azure AI Search \u2013 the proof layer\u00a0<\/h2>\n
\n\u00a0\u00a0“id”:\u00a0“INS-SYN-004-0”,\u00a0
\n\u00a0\u00a0“workspace_id”:\u00a0“ostermere-insurance-demo”,\u00a0
\n\u00a0\u00a0“source_id”:\u00a0“INS-SYN-004”,\u00a0
\n\u00a0\u00a0“title”:\u00a0“Excessive-Wind Zone Underwriting Replace”,\u00a0
\n\u00a0\u00a0“content material”:\u00a0“For brand new\u00a0Hearthmere\u00a0enterprise in zone H3…”,\u00a0
\n\u00a0\u00a0“chunk_number”:\u00a00,\u00a0
\n\u00a0\u00a0“content_vector”:\u00a0[0.012,\u00a0-0.008,\u00a00.031,\u00a0“…”]\u00a0
\n}\u00a0<\/p>\n
\n\u00a0
\nvector_query\u00a0=\u00a0VectorizedQuery(\u00a0
\n\u00a0\u00a0\u00a0 vector=self.mannequin.embed(question),\u00a0
\n\u00a0\u00a0\u00a0\u00a0k_nearest_neighbors=max(top_k,\u00a010),\u00a0
\n\u00a0\u00a0\u00a0 fields=”content_vector”,\u00a0
\n)\u00a0
\n\u00a0
\noutcomes\u00a0=\u00a0self.shopper.search(\u00a0
\n\u00a0\u00a0\u00a0\u00a0search_text=question,\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0# BM25 leg\u00a0
\n\u00a0\u00a0\u00a0\u00a0vector_queries=[vector_query],\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0# vector leg\u00a0
\n\u00a0\u00a0\u00a0\u00a0filter=f”workspace_id\u00a0eq ‘{self.workspace_id}'”,\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0# safety trim\u00a0
\n\u00a0\u00a0\u00a0 choose=[“id”,\u00a0“source_id”,\u00a0“title”,\u00a0“content”,\u00a0“chunk_number”],\u00a0
\n\u00a0\u00a0\u00a0 prime=top_k,\u00a0
\n)\u00a0<\/p>\n![\"\"](\"https:\/\/contributor.insightmediagroup.io\/wp-content\/uploads\/2026\/07\/image-294.png\")
13. Cosmos DB \u2013 the information layer\u00a0<\/h2>\n
\n\u00a0 paths: [‘\/workspace_id’]\u00a0
\n\u00a0 form: ‘Hash’\u00a0
\n}\u00a0<\/p>\n
\n\u00a0 mum or dad: cosmos\u00a0
\n\u00a0 identify:\u00a0guid(cosmos.id, id.id, ‘data-contributor’)\u00a0
\n\u00a0 properties: {\u00a0
\n\u00a0\u00a0\u00a0\u00a0principalId: id.properties.principalId\u00a0
\n\u00a0\u00a0\u00a0 \/\/ 00000000-…-000000000002 is the built-in Cosmos DB Information Contributor\u00a0
\n\u00a0\u00a0\u00a0\u00a0roleDefinitionId: ‘${cosmos.id}\/sqlRoleDefinitions\/00000000-0000-0000-0000-000000000002’\u00a0
\n\u00a0\u00a0\u00a0 scope: cosmos.id\u00a0
\n\u00a0 }\u00a0
\n}\u00a0<\/p>\n![\"\"](\"https:\/\/contributor.insightmediagroup.io\/wp-content\/uploads\/2026\/07\/image-286.png\")
14. Microsoft Foundry \u2013 fashions now, brokers later\u00a0<\/h2>\n
\nfrom\u00a0azure.id\u00a0import\u00a0DefaultAzureCredential, get_bearer_token_provider\u00a0
\n\u00a0
\ncredential\u00a0=\u00a0get_bearer_token_provider(\u00a0
\n\u00a0\u00a0\u00a0\u00a0DefaultAzureCredential(),\u00a0“https:\/\/cognitiveservices.azure.com\/.default”\u00a0
\n)\u00a0
\nshopper\u00a0=\u00a0OpenAI(base_url=”https:\/\/YOUR-RESOURCE.openai.azure.com\/openai\/v1\/”,\u00a0api_key=credential)\u00a0
\n\u00a0
\nresponse\u00a0=\u00a0shopper.responses.create(mannequin=”YOUR-CHAT-DEPLOYMENT”,\u00a0enter=immediate)\u00a0<\/p>\n
\nsearch_wiki()\u00a0
\nget_concept()\u00a0
\ncheck_contradictions()\u00a0\u00a0\u00a0\u00a0\u00a0 \u2190 the gate, as a software\u00a0
\npropose_wiki_patch()\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 \u2190 proposes solely; can’t apply\u00a0
\napply_approved_patch()\u00a0\u00a0\u00a0\u00a0\u00a0 \u2190 requires an approval token\u00a0
\nexport_obsidian_vault()\u00a0<\/p>\n15.\u00a0FastAPI\u00a0on Container Apps\u00a0<\/h2>\n
\nGET\u00a0 \/wiki\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 listing objects, filterable by kind\u00a0
\nGET\u00a0 \/wiki\/contradictions\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 the contradiction register\u00a0
\nGET\u00a0 \/wiki\/{item_id}\u00a0
\nPOST \/question\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 {\u00a0query, mode,\u00a0top_k,\u00a0as_of\u00a0}\u00a0
\nPOST \/ingest\/textual content\u00a0
\nPOST \/ingest\/file\u00a0
\nPOST \/cost-estimate\u00a0
\nPOST \/export\/obsidian\u00a0
\nPOST \/export\/obsidian.zip\u00a0<\/p>\n
\n\u00a0 \/\/ For manufacturing, 1 heat reproduction avoids chilly begins; for a demo, 0 prices nothing.\u00a0
\n\u00a0\u00a0minReplicas: minReplicas\u00a0
\n\u00a0\u00a0maxReplicas: 3\u00a0
\n}\u00a0<\/p>\n16. Id \u2013 no keys, anyplace\u00a0<\/h2>\n
\naz\u00a0deployment group create\u00a0-g\u00a0rg-wikirag-demo\u00a0-f\u00a0infra\/fundamental.bicep\u00a0-p\u00a0namePrefix=wikirag\u00a0<\/p>\n17. The ingestion lifecycle\u00a0<\/h2>\n
![\"\"](\"https:\/\/contributor.insightmediagroup.io\/wp-content\/uploads\/2026\/07\/image-295-1024x760.png\")
\n\u00a0\u00a0\u00a0\u00a0“””Actual id \u2192 title \u2192 alias.\u00a0
\n\u00a0
\n\u00a0\u00a0\u00a0 With out this, ‘money settlement foundation’, ‘depreciated worth’ and ‘ACV’ every turn out to be\u00a0
\n\u00a0\u00a0\u00a0 their very own\u00a0web page\u00a0and the wiki fragments into synonyms. The demo stops at alias\u00a0
\n\u00a0\u00a0\u00a0 matching. A manufacturing system provides embedding similarity and an LLM adjudication\u00a0
\n\u00a0\u00a0\u00a0 step for the ambiguous center, which is the place the fascinating failures dwell.\u00a0
\n\u00a0\u00a0\u00a0 “””\u00a0
\n\u00a0\u00a0\u00a0 candidates\u00a0=\u00a0{identify.decrease(),\u00a0*(a.decrease()\u00a0for\u00a0a\u00a0in\u00a0aliases)}\u00a0
\n\u00a0\u00a0\u00a0\u00a0for\u00a0idea\u00a0in\u00a0self.repository.list_items(“idea”):\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0if\u00a0idea[“id”]\u00a0==\u00a0slugify(identify):\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0return\u00a0idea\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 recognized\u00a0=\u00a0{idea[“title”].decrease(),\u00a0*(a.decrease()\u00a0for\u00a0a\u00a0in\u00a0idea.get(“aliases”, []))}\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0if\u00a0candidates\u00a0&\u00a0recognized:\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0return\u00a0idea\u00a0
\n\u00a0\u00a0\u00a0\u00a0return\u00a0None\u00a0<\/p>\nHalf III \u2013 The demonstration\u00a0<\/h2>\n
18.\u00a0Ostermere\u00a0Mutual\u00a0<\/h2>\n
19. Walkthrough one \u2013 the scoped rule\u00a0<\/h2>\n
\n\u00a0 “query”: “What’s the roof inspection threshold?”,\u00a0
\n\u00a0 “mode”: “hybrid”\u00a0
\n}’\u00a0<\/p>\n
\n\u00a0\u00a0“id”:\u00a0“h3-roof-inspection-threshold”,\u00a0
\n\u00a0\u00a0“kind”:\u00a0“choice”,\u00a0
\n\u00a0\u00a0“abstract”:\u00a0“For brand new\u00a0Hearthmere\u00a0enterprise in zone H3, request a roof inspection when the\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 major roof overlaying is greater than 15 years outdated. Outdoors H3, the overall\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 threshold of greater than 20 years\u00a0stays\u00a0in drive…”,\u00a0
\n\u00a0\u00a0“scope”:\u00a0“New enterprise solely. Zone H3 solely. Doesn’t apply to in-force insurance policies\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 written earlier than the zone mannequin existed.”,\u00a0
\n\u00a0\u00a0“rationale”:\u00a0“Between 15 and 20 years, the frequency of whole or near-total overlaying\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 displacement within the severe-wind band runs at roughly 3.4x the usual\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 band on the similar roof age…”,\u00a0
\n\u00a0\u00a0“rationale_source”:\u00a0“INS-SYN-018”,\u00a0
\n\u00a0\u00a0“accountable_owner”:\u00a0“D. Lindqvist (Head of Property Underwriting)”\u00a0
\n}<\/p>\n20. Walkthrough two \u2013 the contradiction\u00a0<\/h2>\n
\n\u00a0 “query”: “Is hint and entry coated below\u00a0Hearthmere?”,\u00a0
\n\u00a0 “mode”: “wiki”\u00a0
\n}’\u00a0<\/p>\n
\n\u00a0\u00a0“reply”:\u00a0“The place is NOT SETTLED. The information base holds an unresolved\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 contradiction overlaying this query, so no reply is given…”,\u00a0
\n\u00a0\u00a0“warnings”:\u00a0[\u00a0
\n\u00a0\u00a0\u00a0\u00a0“UNRESOLVED CONTRADICTION (con-001): Trace and Access – Standard Cover or Paid\u00a0
\n\u00a0\u00a0\u00a0\u00a0 Endorsement? The system will not choose between the conflicting sources.\u00a0
\n\u00a0\u00a0\u00a0\u00a0 Owner: Y. Tanaka (Product).”\u00a0
\n\u00a0\u00a0],\u00a0
\n\u00a0\u00a0“contradictions”:\u00a0[{\u00a0
\n\u00a0\u00a0\u00a0\u00a0“id”:\u00a0“con-001”,\u00a0
\n\u00a0\u00a0\u00a0\u00a0“status”:\u00a0“unresolved”,\u00a0
\n\u00a0\u00a0\u00a0\u00a0“accountable_owner”:\u00a0“Y. Tanaka (Product)”,\u00a0
\n\u00a0\u00a0\u00a0\u00a0“statements”:\u00a0[\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0{\u00a0“source_id”:\u00a0“INS-SYN-012”,\u00a0“locator”:\u00a0“Chapter 7.3”,\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0“effective_date”:\u00a0“2025-11-01”,\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0“statement”:\u00a0“Trace-and-access costs are covered as standard under\u00a0Hearthmere,\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 up to EUR 5,000…”\u00a0},\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0{\u00a0“source_id”:\u00a0“INS-SYN-008”,\u00a0“locator”:\u00a0“HS-TA-01”,\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0“effective_date”:\u00a0“2026-01-01”,\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0“statement”:\u00a0“Trace and access is an optional endorsement (HS-TA-01)…\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 Where it does not appear on the policy schedule, trace-and-access\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 costs are not payable.”\u00a0}\u00a0
\n\u00a0\u00a0\u00a0\u00a0],\u00a0
\n\u00a0\u00a0\u00a0\u00a0“why_not_resolved”:\u00a0“Each paperwork are present. Neither supersedes the opposite. They\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0 had been written by totally different groups. The Endorsement Catalogue is the more moderen\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0 doc, however recency will not be applicability…”\u00a0
\n\u00a0\u00a0}]\u00a0
\n}\u00a0<\/p>\n21. Walkthrough three \u2013 the date of loss\u00a0<\/h2>\n
\n\u00a0 “query”: “What roof inspection threshold utilized to this property?”,\u00a0
\n\u00a0 “mode”: “proof”,\u00a0
\n\u00a0 “as_of”: “2026-02-20”\u00a0
\n}’\u00a0<\/p>\n
\n\u00a0\u00a0“source_id”:\u00a0“INS-SYN-009”,\u00a0
\n\u00a0\u00a0“in_force_at_as_of”:\u00a0true,\u00a0
\n\u00a0\u00a0“be aware”:\u00a0“outdated at the moment, however in drive on 2025-08-01 – that is the model that\u00a0
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 governs a loss on that date”\u00a0
\n}\u00a0<\/p>\nYet one more,\u00a0dwell\u00a0and unscripted: the multi-hop query\u00a0<\/h3>\n
22. Seeing it: the Obsidian vault\u00a0<\/h2>\n
\n\u00a0\u00a0\u00a0 Residence.md\u00a0
\n\u00a0\u00a0\u00a0 Open Questions.md\u00a0
\n\u00a0\u00a0\u00a0 Ideas\/\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 19 pages\u00a0
\n\u00a0\u00a0\u00a0 Sources\/\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 21 pages\u00a0
\n\u00a0\u00a0\u00a0 Choices\/\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 5 pages\u00a0
\n\u00a0\u00a0\u00a0 Comparisons\/\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 4 pages\u00a0
\n\u00a0\u00a0\u00a0 Contradictions\/\u00a0\u00a0\u00a0\u00a0 2 pages\u00a0\u00a0 \u2190 con-001, con-002\u00a0
\n\u00a0\u00a0\u00a0 Processes\/\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 2 pages\u00a0<\/p>\n![\"\"](\"https:\/\/contributor.insightmediagroup.io\/wp-content\/uploads\/2026\/07\/image-283-1024x340.png\")
![\"\"](\"https:\/\/contributor.insightmediagroup.io\/wp-content\/uploads\/2026\/07\/image-296.png\")
23. The fee argument, actually\u00a0<\/h2>\n
\nEasy RAG\u00a0= Q \u00d7 Tr\u00a0
\nHybrid\u00a0= Q \u00d7 (Tw + V \u00d7 Television)\u00a0<\/p>\n![\"\"](\"https:\/\/contributor.insightmediagroup.io\/wp-content\/uploads\/2026\/07\/image-297-1024x731.png\")
![\"\"](\"https:\/\/contributor.insightmediagroup.io\/wp-content\/uploads\/2026\/07\/image-284.png\")
24. Governance, restated\u00a0<\/h2>\n
25. What I’d construct subsequent\u00a0<\/h2>\n
26. To sum all of it up\u00a0<\/h2>\n
References\u00a0<\/h2>\n