Contents
Get a scoped quote

Enterprise RAG development: Knowledge assistants that cite your documents

For staff or customers asking questions your documents answer: answers built to cite each claim's passage, decline when not covered, and respect source permissions.

Get a scoped quote

First reply within 1 business day

See what you get

Narrated video: it opens on what we build, knowledge assistants designed to cite each claim's passage, decline when nothing covers a question and show each user only what they can open, as of the last sync, then shows The Markup's reporting on New York City's business chatbot, a published study in which about a quarter of citations did not support their sentence, and how Veriprajna checks each claim against its passage on a test set approved by your content owner, using an example on the federal leave regulations where unsupported answers are counted beside declines on questions your documents do answer, before closing on the next step, a free 30-minute discovery call.

Fee
Fixed scope and fee agreed in writing before work starts.
Where it lives
Work products go into your repository as they are produced.
Duration
Building one assistant takes 8 to 12 weeks, including the test set and baseline, fixed in the written scope.

What do RAG development services include?

RAG (retrieval-augmented generation) development builds an assistant that finds the passages in your documents that bear on a question and writes its answer from them. You get a working assistant over your documents, the tests that check it, and the code, all yours.

Read more

The term comes from research that paired a language model with a searchable index of documents (Lewis and colleagues, 2020). In a company, the index holds your policies, contracts, manuals or tickets, and retrieval runs when the question is asked. When a document changes, the next sync changes the answers; no model is retrained.

What makes a knowledge assistant's citations trustworthy?

A citation is trustworthy when its passage supports the sentence, and the assistant is built to decline when none does. We check support per claim on a test set your content owner approves, and count unsupported answers beside declines on answerable questions.

Read more

Why the two counts sit together: an assistant can bring its unsupported answers down by declining more often, including on questions your documents do answer. Counted side by side, that trade shows up instead of passing for progress.

A link on its own proves little until someone reads the passage behind it. Published research on how often cited sentences are actually supported is under Cited is not the same as supported.

Unsupported answers counted side by side with declines on answerable questions: an assistant can bring its unsupported answers down by declining more often, and counted side by side, that trade shows up instead of passing for progress.

The Markup found NYC's chatbot telling businesses to break the law

In 2024, The Markup published its tests of New York City's AI chatbot for business owners and what it said, in a story copublished with Documented and THE CITY.

Source: The Markup, 29 Mar 2024; The Markup and THE CITY, 30 Jan 2026, updated 4 Feb 2026

The Markup also reported that, in practice, it often gave answers without any links. Next: what a cited answer, a decline and a withheld document look like.

Read more

The Markup reported the tips answer as wrong and noted that tips can count toward minimum wage in some cases.

Asked whether buildings must accept Section 8 vouchers, the chatbot said no. The Markup noted that in New York City "it's illegal for landlords to discriminate by source of income, with a minor exception for small buildings where the landlord or their family lives." The Markup also found that similar errors appeared when the questions were asked in other languages.

At the time, a spokesperson for the NYC Office of Technology and Innovation, Leslie Brown, said the city had been clear the chatbot was a pilot program and would improve, and that it "has already provided thousands of people with timely, accurate answers".

In January 2026, Mayor Zohran Mamdani, speaking about a budget gap, called it "an AI chatbot that was functionally unusable". The Markup and THE CITY later reported that the city had taken it down, its web page stating that the "beta test has ended" (The Markup and THE CITY, 30 Jan 2026, updated 4 Feb 2026).

The reporting does not describe how the chatbot retrieved or checked its answers, so this page draws no conclusion about its design.

New York City launched its business chatbot in 2023 as a pilot, announced as drawing on more than 2,000 NYC Business web pages. The Markup found it telling businesses to break the law, for example: "Yes, you can take a cut of your worker's tips." The city has since taken it down.

The chatbot's tips answer as The Markup published it, beside The Markup's note that the answer was wrong and that tips can count toward minimum wage in some cases. Below it, a policy question, a confident answer, and an empty space where its source should be: no source to check.

What you get: Cited answers, declines and permission tests

The rows are an example written on a real federal regulation. Each part we deliver points to the row it produces.

  • Your report states its test set, its n and the unit of each count; the test set is frozen before the build phase, so results compare like for like.
  • A rate on real questions comes only from a random sample of your own questions, and questions outside the test set can still be answered wrongly.
  • Your test set and results are yours; what you tell your customers about your own testing is your decision. The report is prepared for you only; no third party may rely on it, and we issue no letters, and we do not name ourselves to your customers without written consent.
  • The report is test evidence for your own release decisions, not a certification.
  • Cited answers: Each factual sentence linked to its passage

    Each citation shows the passage itself, with its document and section, so a reader can check the sentence against it without leaving the answer.

  • Decline rule: No supporting passage, no answer

    Where a declined question is sent, such as a team inbox or a named page, is set with you in the written scope.

  • Permission filtering in the retrieval query, per source in scope, synced on schedule

    Per-role tests run with test users and are reported as passed of run, per role, with what was withheld. A clean result means no restricted passage was returned in the tests that ran, not that none can be. Revoked access takes effect at the next sync. Permissions that are too broad in a source stay too broad in the assistant; the assistant does not fix them.

  • Test suite and test set in your repository, set to run in your CI

    We draft the test set from your documents; your content owner approves it before the baseline runs; the approved set and its labels are yours.

    We label; your content owner reviews a stated sample, including every case where the automated judge disagreed; the report shows how many labels the owner changed.

    The written scope fixes how many items we label by hand; the automated judge only extends the check beyond them. If the judge falls short of the agreement level set in the written scope, results are reported on the hand-labelled items alone, and the fee is unchanged.

    Your own questions, if you want them: your content owner writes a few from real cases never sent to us and keeps them from us; your engineer runs them at the baseline and again at handover. They print as their own row, run by you; we see counts only, and the scripts and judge are ours. If you decline, the report says the whole test set was visible to those building the assistant.

    An answer counts as unsupported if any factual sentence in it is unsupported; the report gives both counts, per claim and per answer, with their n.

    If a threshold is missed, the one fix round on the change list is within the fixed fee, and the re-run uses the frozen test set.

  • Report

    The test set and its n, the unsupported-answer and decline counts, the permission tests and the change list, with the gate result at handover. Each result says how it was observed: seen by us, on a copy from your export, or confirmed by you.

    Its cover states who drafted and ran the tests. New build: "The tests in this report were drafted by Veriprajna and run as set out in the written scope; Veriprajna also built the assistant tested, so this is not independent assurance." Improvement build: "Veriprajna changed the assistant tested, so this is not independent assurance." Stop after the baseline: "The tests were drafted by Veriprajna and run as set out in the written scope; this is not independent assurance."

  • Also in your repository: Chat page or integration, deployment configuration, ingestion and source runbook
    • The chat page or integration your users ask through, agreed in the written scope
    • Deployment configuration for your cloud accounts, deployed by your CI
    • Ingestion with re-index at each sync
    • Runbook for adding a source
    • Handover session

    Connector changes are quoted in writing, or covered by upkeep where the written scope includes them.

    Everything delivered is yours to keep and change: no runtime call to our systems and no licence fee. The private tooling we use to build it stays ours and is not delivered.

Worked example: One question set on the FMLA regulations

Source: eCFR, 29 CFR Part 825

Question

"I joined this year. Can I take FMLA leave?"

Answer

Three conditions apply: at least 12 months with the employer, which need not be consecutive, though periods before a break in service of seven years or more need not be counted, with exceptions; at least 1,250 hours of service in the 12 months before the leave; and a worksite with 50 or more employees within 75 miles. [1] [2]

[1] § 825.110(a): "(1) Has been employed by the employer for at least 12 months, and (2) Has been employed for at least 1,250 hours of service during the 12-month period immediately preceding the commencement of the leave ... and (3) Is employed at a worksite where 50 or more employees are employed by the employer within 75 miles of that worksite."

[2] § 825.110(b): "The 12 months an employee must have been employed by the employer need not be consecutive months, provided (1) Subject to the exceptions provided in paragraph (b)(2) of this section, employment periods prior to a break in service of seven years or more need not be counted ..."

Not covered

"What is our remote work policy?" No document in scope covers it, so the assistant declines and points the user to the HR team.

Test user without access

A test user in a contractor role asks about a file in a folder that role cannot open. No passage from that folder is retrieved, so nothing from it can be cited, and the assistant declines.

Read more

Kept apart in every report. (A) The written test set: questions drafted from your documents, questions they do not answer, per-role permission tests and injected-instruction cases.

(B) Only where you already have real questions: a random sample of them, drawn as set out in the written scope and requested only as a pseudonymised export. Only (B) is reported as a rate on real questions, with any interval given per question. A set of questions you choose is reported as its own group and never called a rate.

How an engagement runs

Start with a test set and baseline on a sample of your documents, or go straight to a build. Upkeep after handover is optional.

Test set and baseline (optional)New builds: a first pipeline on a document sample, test set, answer and decline counts, permission tests where your sample has them, change list.2 to 3 weeks
Details

The first pipeline runs in our development environment on the exported sample, through the model provider named in the written scope, and its code is committed to your repository as it is produced. Until deletion, the first unit's sample index lives in our development environment.

On an existing assistant, your assistant produces the baseline answers.

BuildDefaultOne assistant over the sources agreed in the written scope; its first phase is the test set and baseline, then the build runs to handover.8 to 12 weeks
Details

On an existing assistant, this is an improvement build: the baseline is measured on your assistant, then the changes are made and re-run on the frozen test set. Where your assistant sits on a platform that cannot be changed, its test set carries over to a new build. If we validated or independently assured your assistant, we do no build or fix work on it for 12 months.

Upkeep (optional)Monthly re-runs of the test suite, new sources and a review of declines, as set in the written scope.Monthly, per scope
  • On an existing assistant, the written scope prices the baseline and change phases separately, and you may stop after the baseline.
  • Expect 2 to 4 hours a week from a content owner and about a day a week from one engineer; fixed in writing.
  • If you bought the test set and baseline first, the build's own length is fixed in the written scope.
  • Fees are staged in the written scope, so halting costs only the stages reached.
  • Documents owned by your client need their written consent to our access and to each model or AI tool, in the written scope.
Read more

Your people provide a content owner who knows the documents and reviews a sample of our labels, the administrators of your sources for the permission export and test users, and an engineer who merges our pull requests and runs the connectors, permission sync, production index, deployment and test runs.

Is this for you?

A fit

  • Answers must come from policies, contracts, manuals or regulations that change, and each answer must cite them.
  • A wrong answer has consequences: policy, legal, regulatory or public information, including a municipal chatbot over official pages.
  • Some content is restricted, and each user may see only what they can open in the source system.
  • You already run an assistant that guesses, and you want it changed, not only measured.

Not a fit

  • Measure an existing assistant without changing it, or build a standalone evaluation suite: AI Hallucination & Output Verification Audit or LLM & Agent Evaluation.
  • Questions answered from databases or warehouses: Conversational Analytics & Text-to-SQL.
  • An agent that sends, writes or buys on your behalf: AI Agent Development.
  • Model hosting on your own hardware or a closed network: Private & Sovereign LLM Deployment.

Choosing a RAG development company: How we work with you

What we commit to, and how your documents, permissions and systems stay protected while we build.

  • A mutual NDA and the written scope are signed before we read anything.
  • Your engineer first scans the repository and exports for keys and unneeded personal data; anything we still find is reported and not used.
  • No role that can push until production secrets are out of its reach.
  • Our changes arrive as pull requests you merge, through credentials you issue and revoke, without production secrets or deploy rights.
  • Documents, permission data and questions reach us only as listed exports; real questions only pseudonymised.
  • Our copies are deleted on request, or 30 days after our last included work; contract and billing records stay.
  • Our model calls run on our account, or on a capped non-production key you issue and revoke.
  • We never name you, reference the engagement or publish findings without your written consent.
Read each control in full
  1. 1. A mutual NDA and the written scope are signed before we read your code or documents.
  2. 2. Before we read anything, your engineer scans the repository, git history included, and each export for keys and for personal data the scope does not call for. Anything we still find is reported for you to rotate or remove; we do not use it, pause on that file until cleaned, and name any agreed AI tool that already processed it.
  3. 3. Before we get a role that can push, your engineer protects the main branch and moves production secrets, CI secrets included, out of that role's reach, or gives us a fork or a separate repository. Where a branch cannot be protected, our work goes into a separate private repository you own.
  4. 4. Changes arrive as branches and pull requests you review and merge. Credentials you issue reach your repository and CI without production secrets or deploy rights, as listed in the written scope; you revoke them at handover and issue fresh ones if a re-test fix needs them. The assistant runs in your cloud accounts, deployed by your CI.
  5. 5. We work from what you send: document samples, unredacted only where answers need the text; a permission export and test-user definitions; test questions and test-run results; real questions only as a pseudonymised export. Anything identifying that we did not ask for is reported to you and not used.
  6. 6. Code, document samples, exports, the sample index and test material are deleted from our systems on request, or 30 days after our last included work: the re-test window, the baseline if you stop there, or upkeep's last month. Contract and billing records stay.
  7. 7. Model calls during our work, including the first pipeline on your sample, use our account with the model provider named in the written scope, within a budget we carry. If you prefer your own account, you issue a separate non-production key with a spend limit and revoke it afterwards.
  8. 8. We never name you, reference the engagement or publish findings without your written consent.

Access drawing. On your side: your cloud accounts, where the assistant runs; your sources and production index, which your engineers connect and sync; your repository, where our pull requests, made through credentials you issue, wait for your merge; and CI without production secrets or deploy rights. Crossing to us: document samples, unredacted only where answers need the text; a permission export and test-user definitions; test questions and test-run results; and real questions only as a pseudonymised export. On our side, until deletion: copies of what you send and the first unit's sample index in our development environment.

You get

  • Go-live thresholds set with you in writing after the approved test set and baseline counts
  • The gate result on the frozen test set at handover, on the build's last day
  • If a gate threshold is missed, one fix round on the change list is included, then a re-run
  • The full report within 5 business days of the build ending
  • The baseline report within 5 business days of the baseline ending, when you stop there
  • One re-test within 30 days of handover, including fixing defects in what we delivered
  • Liability capped at the fees paid for the engagement
  • The test set and baseline is an optional way to try us; you keep its test set, counts and change list whether or not the build follows; the build can always be quoted directly.

A platform's built-in assistant, or a tested build

With a platform's assistant alone, these checks stay with your team. Here is what we deliver for each, fixed in the written scope.

QuestionLeft to your teamWhat we deliver

Your team decides how to check each sentence against its passage, and runs the check.

Each sentence checked against its passage on the approved test set, with declines on questions it could answer counted too.

Your team sets what the assistant should say, and tests that it says it.

A decline rule, with declined questions sent to the inbox or page you choose.

Your team writes per-role tests and checks what each role gets back.

Access filtering inside the retrieval query, per source, with each role's results in the report.

Your team sets thresholds, builds the test set and decides what a miss means.

Thresholds agreed in writing once the baseline counts are in, and one included fix round on a miss.

Your team connects and keeps in sync any source the platform does not reach.

Connectors and permission sync for each agreed source, each checked before build work on it.

Cited is not the same as supported

A citation shows where an answer says it came from. Whether that passage says what the sentence says is a separate question, and published research has measured the gap.

74.5%

of citations supported their sentence, on average, across four generative search engines tested in 2023 Liu, Zhang and Liang, Findings of EMNLP 2023

Cited, not supported

  • A sentence with a citation whose passage says something else

Cited and supported

  • A sentence whose cited passage says what the sentence says

How we count support and declines: the Tests step in How we build it.

Read more

Liu, Zhang and Liang at Stanford University had human annotators check the answers of four public generative search engines to 1,450 queries, with responses collected between late February and late March 2023. The figure above is citation precision. The same study found that, on average, only 51.5% of generated sentences were fully supported by their citations (citation recall).

These were web search systems, not assistants over a company's own documents. What the study shows is that a citation's presence does not show that its passage supports the sentence; it does not predict how often your assistant would err. That is what your own test set measures.

How we build it

Six build steps, delivered through pull requests you merge. Select one to see where it sits in the assistant.

The assistant: your sources feed ingestion, which writes passages with their access lists to the index; a user's question goes to retrieval, filtered by that user's access as of the last sync; the answer step writes cited sentences or declines; the test suite, with injected-instruction cases planted only in test copies of documents, runs against the whole path, and your CI compares each run with the frozen baseline. Each build step highlights its part.

Tap a build step to zoom the map to it.

Build steps

  1. SourcesDocuments are parsed with their structure, split into passages that keep their headings, and re-indexed at the next sync after a source changes.

    Each passage keeps its document, section and location, so a citation can show exactly where it came from. Content that changes faster than the agreed sync is a poor fit: answers would trail the source.

    Text inside a document reaches the model as retrieved text. OWASP's Prompt Injection entry lists "Indirect injection through retrieved content" among its examples, which is why the test set carries injected-instruction cases (see Tests).

  2. PermissionsEach passage carries its source's access list, so a user's query is built to reach only passages that list allows, as of the last sync.

    The filter sits inside the retrieval query, not after it, which is where OWASP's Sensitive Information Disclosure and Vector and Embedding Weaknesses entries place it (quoted under Standards).

  3. RetrievalHybrid keyword and semantic search picks the passages; reranking is added only where the test set shows it helps.

    Keyword search finds exact terms such as section numbers, product codes and defined terms; semantic search finds passages worded differently from the question.

  4. AnswersEach sentence is built to link the passage it was written from; a decline tells the user where to ask instead.

    The answer rules are prompts and checks kept in your repository. When the retrieved passages do not support an answer, the rule is to decline rather than fill the gap from the model's general knowledge.

  5. TestsThe support judge is checked against human labels on a held-out split; declines are counted beside unsupported answers, so more declines show in the counts.

    The test set holds questions your documents answer, questions they do not, per-role permission tests with test users, and injected-instruction cases planted only in test copies of documents. Adversarial testing beyond those cases is AI Red Teaming & LLM Penetration Testing.

    Documents and questions in the languages agreed in the written scope; the test set and labels are drafted in those languages, and the content owner's review covers each. Where a language needs a native-speaker rater, that rater is your own reviewer or one named in the written scope, bound by confidentiality before access.

    The judge in the delivered suite runs through the model provider named in the written scope; any judge we use only while building is kept separate from it.

  6. RunningEach run reports its counts against the frozen baseline, so every release is compared like for like.

    Your CI runs the suite before each release, on your non-production index. Monthly re-runs, new sources and decline reviews come with Upkeep.

Where the build meets OWASP and NIST

Each framework below names a risk that a build step addresses. How far that goes depends on your system and how you run it.

Answers

One build step, mapped to the frameworks below.

  • OWASP Top 10 for LLM Applications 2026Misinformation
  • NIST AI 600-1, Generative AI Profile (July 2024)Section 2.2, Confabulation
  • OWASP wording is quoted from "OWASP Top 10 for LLM Applications 2026", Version 2026. Mapping a build step to these frameworks is our reading of them, not a compliance finding.
  • Other steps: Permissions maps to Sensitive Information Disclosure and Vector and Embedding Weaknesses; Sources to Prompt Injection.
Read more

OWASP, Sensitive Information Disclosure. Its foundational tier, marked "every deployment", says: "Authorize before retrieval: enforce document- and chunk-level authorization inside the index query, not at the application layer after retrieval." It also names oversharing: "unscoped drives, legacy permissions, and knowledge bases feed RAG with sensitive data the model then retrieves as designed. The fix is the data surface, not the model."

OWASP, Vector and Embedding Weaknesses. "Enforce tenant scoping inside the index query, not as a post-retrieval filter, and validate it server-side." And: "Apply access control at the chunk level: a mostly-public document can contain a confidential paragraph."

OWASP, Prompt Injection. "Indirect injection through retrieved content: attacker instructions ride in a RAG passage, web page, document, or email and run when the content reaches the context window."

OWASP, Misinformation. Misinformation "occurs when an LLM or LLM-enabled application produces incorrect, incomplete, unsupported, or misleading information that appears credible enough to influence a human decision, an automated workflow, or an agent action."

NIST AI 600-1. Confabulation is "the production of confidently stated but erroneous or false content (known colloquially as "hallucinations" or "fabrications") by which users may be misled or deceived."

Questions buyers ask

7 questions

Retrieval, for knowledge that changes: a passage such as § 825.110(a) reaches the model at answer time and the answer is built to cite it or to decline, so the passage is in front of the reader. Fine-tuning, for style and behaviour: tone, format and domain vocabulary change the model itself; it changes how a model writes or behaves. The two can be combined.

Fine-tuning or RAG: Which fits our documents?

For answers that must cite your documents and follow them as they change, RAG: retrieval puts the source passage in front of the model and the reader at answer time, so each sentence can be checked against it. Fine-tuning changes how a model writes or behaves, and it can be combined with retrieval. When style or behaviour is the goal, see Fine-Tuning & Domain Language Models.

Should we use a platform's built-in assistant instead?

Sometimes that is enough: Microsoft 365 Copilot, Glean and Amazon Bedrock Knowledge Bases each answer questions over company content. What a build adds is set out under A platform's built-in assistant, or a tested build. We are not affiliated with any company or product named on this page, and take no commission from any model or tool vendor.

Will users see documents they are not allowed to open?

The assistant is built to return only passages the user can open in the source system, as of the last permission sync. Fixing overshared permissions in the sources themselves is separate work: AI Data Security & Copilot Readiness.

How accurate will it be, and what if it gives a wrong answer after launch?

No figure is promised in advance, and questions outside the test set can still be answered wrongly. Go-live is your decision. The gate result is reported at handover, pass or fail; the gate is not a payment or acceptance condition unless the SOW says so. After the included re-run, further changes are quoted in writing. Liability is capped at the fees paid for the engagement.

What does a RAG build cost?

The fixed fee follows your sources, document volume and formats, permission model, languages and model choice. Running cost follows query volume and the model, search and hosting services, billed to your accounts; there is no licence fee and no runtime call to our systems. An existing assistant's test set carries over to a new build. You get a written scope and fixed fee within 2 business days of a free 30-minute call.

When can you start?

The earliest start date is in the written scope. The fixed fee does not change if the schedule slips on our side, and you can halt or end the engagement at any time.

Can you build on our sources and our model provider?

Connectors and permission sync are written for the sources agreed in the written scope; each source's permission model is checked in the baseline, before any build work on it. If the check finds a source the written scope did not assume, the change is quoted in writing before work on it. At answer time the assistant calls the model provider named in the written scope, through your account.

Misgrounded answers, fine-tuning, agents and graphs

Agentic RAG, CAG, MCP and knowledge graphs, each against the retrieval step. Agentic RAG: the model decides when to search again, rewrite its query or pick a different source, used when the test set shows that a single pass misses passages. CAG: all the relevant documents preloaded into the model, so per-user filtering has no retrieval step to sit in. MCP: an MCP server is a way to connect retrieval, not a replacement for it. Knowledge graphs: when questions turn on relationships, entities and links can be retrieved alongside passages.

When a cited answer is still wrong

Not every wrong answer is a made-up one. In their study of AI legal research tools, Magesh, Surani, Dahl, Suzgun, Manning and Ho (Journal of Empirical Legal Studies, 2025) count a response as hallucinated "if it is either incorrect or misgrounded". A response is misgrounded "if key factual propositions are cited but misinterpret the source or reference an inapplicable source."

The authors call these errors "potentially more dangerous than fabricating a case outright, because they are subtler and more difficult to spot." A misgrounded answer looks finished: it has a citation, and the citation is real. Only reading the passage shows the problem.

That is why the test suite checks support per claim, sentence against passage, rather than checking that a citation exists. The same gap between cited and supported is measured in the study under Cited is not the same as supported.

Fine-tuning or retrieval: When each fits

The FAQ gives the short answer. The longer one turns on what has to change.

  • Knowledge that changes: policies, prices, contracts, product manuals. Retrieval reads the current text at question time, so an update reaches answers at the next sync, with no retraining.
  • Style and behaviour: tone, format, domain vocabulary, how the model handles a task. That is what fine-tuning changes.
  • Both: a fine-tuned model can still answer from retrieved passages, and cite them.

Citation is not exclusive to retrieval: Khalifa and colleagues trained models to cite the training document that supports an answer, in experiments on synthetic data. For this service the point is narrower. With retrieval, the passage sits in front of the reader at answer time, and the test suite can check each sentence against it.

When the goal is a model that writes or behaves differently, Fine-Tuning & Domain Language Models covers it.

Agentic RAG, CAG, MCP and knowledge graphs

Agentic RAG. The model decides when to search again, rewrite its query or pick a different source, instead of retrieving once. It is a design choice inside a build, used when the test set shows that a single pass misses passages. An agent that sends, writes or buys on your behalf is AI Agent Development.

CAG. Cache-augmented generation, as proposed by Chan and colleagues, preloads all the relevant documents into a long-context model and answers "without additional retrieval steps". It is aimed at documents "of a limited and manageable size". With everything preloaded, per-user permission filtering has no retrieval step to sit in, so restricted content needs a separate preloaded set per permission group.

MCP. The Model Context Protocol is "a standard for connecting AI applications to external systems", such as data sources and tools. An MCP server can expose search over your documents to an assistant, so MCP is a way to connect retrieval, not a replacement for it. Building MCP servers is MCP Server Development.

Knowledge graphs. When questions turn on relationships, such as which contract governs which supplier, or which policy replaced which, a graph of entities and links can be retrieved alongside passages. If the test set shows passage retrieval missing those questions, Knowledge Graph & Context Engineering covers it.

Put a knowledge assistant that cites or declines in front of your users

Building a new assistant, improving one you have, or scoping one for a tender or a client: start with a short conversation about your documents and who asks them.

  1. Email or WhatsApp us
  2. Reply within 1 business day
  3. Free 30-minute discovery call
  4. Written scope and fixed fee within 2 business days of the call