Axiom — enterprise retrieval, not just another chat wrapper.
A retrieval-augmented writing tool for engineering teams. We built the retrieval, eval
harness, and production pipeline. The demo below reconstructs that pipeline with sample
questions and pre-baked answers — pick one and watch the flow.
ClientAxiom AI, Inc.
SectorApplied AI · B2B
Engagement11 months
Team4 eng · 1 applied-ML
StackPython · pgvector · bge-large · Ray
Shipped2025 Q3
§ Interactive demo · RAG pipeline
Ask a question. Watch the retrieval, watch the answer.
❯or pick:
§ RETRIEVED · TOP 4similarity
docs/citations.md#disambig0.91
Axiom disambiguates citations using a three-stage ranker: lexical overlap, entity-type filter, then cross-encoder rerank.
rfcs/RFC-023.md#rerank0.88
The rerank uses a distilled bge-large model at 76ms p95. If confidence is below 0.55, we prompt the author for clarification.
runbooks/citations.md0.74
On-call: if the disambiguator is failing, fall back to "lexical only" by flipping the AXIOM_DISAMBIG_MODE=lex flag.
tests/disambig.json0.69
Gold set of 1,240 ambiguous citation queries. Current pass rate: 93.6%.
§ GENERATED · STREAMINGmodel: bge-large-v3
RETRIEVED
0chunks
P95 LATENCY
0ms
TOKENS
0
COST
$0
§ What we shipped
An eval harness, a retrieval pipeline, and a production runbook.