Service 02
RAG Assistants & AI Agents
An assistant that has actually read your files, and an agent that can act on what it finds — with retrieval built so the answer is traceable back to the document it came from.
Teams sitting on a body of private knowledge — precedents, policies, contracts, past matters — that is technically searchable and practically unsearchable.
retrieve → reason → act
General models know everything except the thing you need: what is in your files.
Naive retrieval returns plausible passages rather than correct ones, and a confident wrong answer is worse than no answer at all.
Once more than one client's data is involved, retrieval has to prove it cannot reach across the boundary — filtering results after the fact is not an answer.
How it's built
Five stages, and a human at the end.
Each stage is something you can watch run. Nothing here is a black box you're asked to take on faith.
Ingest
Documents are parsed, chunked with structure preserved, embedded, and written to a vector store scoped to the tenant or matter they belong to.
Isolate
Access is enforced at the storage layer, not by post-filtering — the same pattern I worked out in the open-source multi-tenant RAG build.
Retrieve
Queries are expanded and reranked so the passages that reach the model are the ones that answer the question, not merely the ones that share its vocabulary.
Ground
Answers cite the source passage. If the retrieval is weak, the assistant says so instead of filling the gap with fluent invention.
Act
Where it's warranted, the assistant becomes an agent: updating the CRM, triggering a workflow, opening a task — with the boundaries of what it may do written down and enforced.
- A document ingestion pipeline you can re-run
- A vector store with per-tenant or per-matter isolation
- A retrieval layer tuned and evaluated on your own questions
- Citations on every answer
- Defined action boundaries for anything agentic
- An interface your team will actually open — chat, or inside a tool they already use
No client case study is published for this service yet. The engineering behind it is public, though — the multi-tenant retrieval work is open source and you can read the code.
See the buildDocument & Intake Automation
Losing hours to work an AI system could handle?
Tell me about the process. I'll tell you straight whether it's worth automating — no pitch, no obligation.
Book a call