Answer from your own docs with an LLM agent
October 31, 2026 · release
@voicethere/agent 0.21.0
Our landing page assistant used to answer from three markdown files that someone had to keep in step with the real docs. Now it searches the rendered docs pages themselves, so when a price changes on a page, the next build changes the answer.
The part that does the searching is not specific to us. We published it as an image, docs-rag-mcp, together with an llm-docs-rag template in @voicethere/agent 0.21.0, so you can point an LLM agent at your own documentation.
How it fits together
You build your docs into the image with one command, and the build can fail when retrieval gets worse. The image runs as a project service next to your agent. The agent lists its two tools, search_docs and get_doc, and the template's prompt tells the model to search before it states a fact and to name the page it found it on.
Search combines meaning and keywords: an embedding model that runs inside the container finds the right passage for a loosely worded question, and full-text search catches exact numbers and names such as 45 credits per hour. There is no embedding API key and no outside call per question, and nothing is written at runtime.
Try it
- Guide: Docs knowledge base (RAG)
- Template and image source: llm-docs-rag and docs-rag-mcp
- Related: LLM agents and Project service
voicethere projects services set docs \
--image registry.example.com/acme/docs-mcp:1 \
--port mcp:8080 --health http:/healthz:8080 \
--run-mode always_on --cpu 500 --memory 1024
voicethere deploy --wait
The model is English-first, so run the eval on your own questions before you rely on another language. The guide lists the limits.