RAG for software documentation: assistants that consult manuals and versions
Published on 2026-11-16 by ArduBot
Design a RAG assistant for technical documentation with version filters, code snippets, sources, and response evaluation.
The user finds a tutorial for the previous version and applies a command that no longer works. In technical documentation, providing a real source is not enough: that source must correspond to the version the person is using.
The documentation problem with multiple versions
Imagine an API that keeps references for v1, v2, and v3. Semantic search can find the v1 example even if the question mentions v3, because both explain the same operation. The solution starts before the model: store product, version, language, updated_at and a permanent link per section.
When the user asks “how do I authenticate requests?”, the assistant should ask for the version if it is not indicated or apply the active version of their project when that information is authorized and available.
What to index and how to preserve code
Separate prose and code examples without breaking their relationship. A function should not lose its language, SDK version, or required parameters. For an API reference, preserving headers, HTTP method, path, fields, and exceptions usually helps more than splitting by fixed length.
Lexical search is useful for identifying exact paths, function names, and error codes. You can combine it with embeddings for questions phrased in natural language, without assuming that a vector perfectly recognizes get_user versus get_users.
Respond with verifiable sources
A useful response indicates the procedure and links to the relevant section and version. If there is a change between versions, it should only be explained when there is evidence. In questions such as “is this method deprecated?”, it is especially important to check dates and documentation notices.
When there is not enough evidence, the assistant can offer related results instead of improvising a command.
Integration with development tools
A RAG API can feed a portal's search engine, a support chat, and an agent that works in the IDE. MCP could expose a document reading tool to compatible agents; it is not necessary to allow actions on production to consult documentation.
It is advisable to strictly separate knowledge queries from deployment operations, credential modification, and service restarts.
How to test the assistant
Prepare questions by version, examples that share method names, and frequent errors. Verify that each citation opens the corresponding version. Add questions requesting functionalities that do not yet exist: a correct response acknowledges that they are not documented.
Measure response time, abandoned searches, tickets that end up being escalated, and corrections due to incorrect responses. A reduction in tickets should be compared with a comparable base period and with a sufficient sample.
What could be sold
For small teams and agencies, a managed knowledge base can save the effort of building separate indexers for each documentation. ArduMaker is evaluating that opportunity; we do not yet offer an automatic RAG node as an active product. If you maintain SaaS documentation, explain how many versions and projects you would like to query from ArduMaker.
References: RAG and hybrid search in Azure · ArduMaker MCP Tools.