Machines / RT-02
MODEL RT-02
Retrieval engine
Your documents in. Cited answers out.
Search over your own documents and data that cites every claim — plus an evaluation harness that flags the moment quality drops.
In plain words: A chatbot that answers from your documents — with sources.
INPUTYour documents and a question
OUTPUTAn answer that cites its sources
RUNS ONYour infra; self-hosted open models when data can't leave
OPERATORA person owns the eval set and approves releases
Operating manual
Rev. 1 · RT-02§1
What it does
Splits your documents into retrievable pieces, finds the right ones for each question, writes an answer that sticks to them, and shows the source. A test set runs on every change so regressions fail in CI, not in front of users.
§2
Install when
- People keep asking the same questions of the same documents
- Answers must be traceable to a source
- Content changes often and the bot must keep up
§3
Do not install when
- The knowledge isn't written down anywhere
- Questions need judgement, not lookup
- Nobody will own the test set
§4 · The line
Each station is a separate step with one job and its own check. A failure is caught where it happens, not at the end.
01QuestionFrom chat, WhatsApp or an app
02SearchHybrid keyword + vector retrieval
03Re-rankBest passages first
04AnswerGrounded draft, refuses if unsure
05CiteSources attached, logged
↺ Eval set runs on every index, prompt or model change → below threshold blocks the release
§5 · Safety features
Says "I don't know"No source, no answer — never a confident guess.
Eval gateRecall, faithfulness and correctness scored on every change.
Permissions respectedUsers only get answers from documents they can see.
Full traceEvery answer logs its sources for audit.
§6 · Gauges
What's watched on every install. Calibrated per team.Retrieval recallWas the right passage found?
FaithfulnessDid the answer stick to it?
CorrectnessDid the user get what they needed?
Cost per answerTokens and compute per question
§7 · Parts bin
Embedding model
Vector store — pgvector / similar
Re-ranker
LLM — hosted or self-hosted
Eval harness in CI