RAG development services for assistants that cite sources

Retrieval-augmented generation (RAG) means the system first searches your documents for passages relevant to a question, then asks a language model to answer from those passages, with citations. The approach was set out by Lewis et al. in 2020 (opens in a new tab) and is now the standard pattern for company knowledge assistants.

For knowledge that changes, RAG beats fine-tuning. Fine-tuning adjusts a model's weights with examples. It is good at teaching a format or a tone and poor at teaching facts that will be stale next quarter. It also cannot show where an answer came from, or hide one department's files from another.

Off-the-shelf tools such as Microsoft 365 Copilot or Glean are a fair first step when your documents sit in systems they already connect to. A custom build pays off when sources fall outside their connectors, when answers must cite the exact passage, or when you need control over how the assistant searches and refuses.

No language model can be made incapable of error, so we design for 3 things: find the right passage, answer only from it, and refuse when it is not there.

Most wrong answers are retrieval failures. The model was handed the wrong paragraphs and did its best with them, so most of our effort goes into search. Keyword search catches part numbers, policy codes and names. Vector search catches meaning when the wording differs. A reranker then reorders the top results. Anthropic's benchmark of this approach found that adding context to each chunk plus keyword search cut failed retrievals by 49%, and reranking took the cut to 67% (Anthropic, Contextual Retrieval, 2024 (opens in a new tab)). Your documents will behave differently, so we measure on your own questions.

The assistant must cite a passage for every claim. If retrieval finds nothing relevant, it says so and names the team that owns the topic. A confident wrong answer does more damage than no answer at all.

The evaluation set scores both halves separately. Did the right passage reach the top results? Did the answer match what your expert approved? When a score drops, we know which half broke. Our glossary entry on AI hallucination explains the failure mode itself.

Yes, and it has to be designed in from week 1, because retrofitting it is expensive. The rule we build to: the assistant never retrieves a passage the person asking could not open in the source system.

That takes 3 pieces of work. Each chunk in the index carries the access list of its source file, synced from SharePoint, Google Drive, Confluence or your document system. Every search is filtered by the user's identity before results reach the model, so a restricted passage never enters the prompt. Filtering the answer afterward is too late. And the sync job updates permissions when they change at the source, so someone removed from the HR folder in the morning loses access in the assistant the same day.

Sign-in uses your identity provider, such as Microsoft Entra ID, Google Workspace or Okta. Every question, the passages retrieved and the answer given are logged for audit.

Permission sync is the biggest single reason multi-source builds land in the $35,000 to $75,000 band. SharePoint passes permissions down from sites to libraries to folders, with exceptions at any level. Google Drive mixes shared drives, groups and sharing links. We map each source during Map, and if its permissions are too tangled to trust, we recommend leaving it out of the first release.

A knowledge assistant over one main source for one audience costs $15,000 to $35,000 to build. Assistants that draw on several systems with permission sync cost $35,000 to $75,000.

The build price moves with 5 things:

• Sources. 2 SharePoint sites share one connector. SharePoint, Confluence and a ticket archive need 3.
• Document condition. Clean Word files are easy. Scanned PDFs, tables split across pages and slide decks that carry their meaning in diagrams need extra parsing.
• Permissions. One audience with shared access is simple. Per-user access synced from several systems is the expensive part.
• Where it lives. A Slack or Teams app ships faster than a custom intranet page or a branded customer widget.
• Expert time. Your subject experts approve the test answers. If they are short of time, the schedule stretches.

Running costs pass through at cost, with no markup: model usage per question, embeddings when documents change and hosting for the index. Index hosting is a fixed monthly cost, and model usage tracks how many questions people ask. We measure both during Prove and put a monthly estimate in the proposal. Monitoring and retesting after model updates are available from $1,500 a month.

RAG answers questions from text. It is the wrong tool for numbers that live in a database. "How many orders shipped late last month?" is a query against your order system, which we build as a read-only database tool for an AI agent, or answer with the dashboard you already have.

It is also the wrong choice when the documents are the problem. If your policies contradict each other, an assistant will quote both with confidence. Then the first project is a documentation cleanup, and we will say so in Map, before you pay for a build. Our guide to building a RAG knowledge assistant covers that call in more detail.

And it may be more than you need. If you are a small team and the answers live in one well-kept Notion workspace or Google Drive, try the AI features built into those tools for a month first.

A knowledge system build does not include:

• Writing or rewriting your documentation. We report gaps and conflicts, and your content owners fix them.
• Migrating files between systems, such as a file server move.
• Legal sign-off on what the assistant may tell customers.
• Software licenses and cloud accounts, which you hold directly.

If the real job is pulling fields out of invoices or contracts, that is intelligent document processing, a different build.