lexicon
Grounded document Q&A that gives cited answers or refuses.
~/problem
Document Q&A that sounds confident but is wrong is costly. A feasibility spike showed similarity scores could not separate a correct answer from a topically adjacent wrong one, so refusal is enforced by a separate verification step.
A FastAPI and Next.js system that answers questions only when the answer is backed by citations in the supplied documents. An independent groundedness verification call, not a similarity threshold, decides between answering and refusing. v1.0.0 is tagged.
~/highlights
- Hybrid retrieval over PostgreSQL with pgvector; retrieval recall@k is measured for real (9/9) because it never calls an LLM.
- Verification call hardened against prompt injection inside passages: 18 of 18 structural containment across an 18-document adversarial corpus.
- Tamper-evident audit trail (ADR-0002) and instance-level authentication (ADR-0005).
- Committed, CI-gated evaluation harness with a golden dataset.
- Production-shaped compose stack with non-root images and a pre-warmed embedding model; three bugs found and fixed by running it.
- No LLM API key exists in the project by design (ADR-0004): all quality numbers measure the harness against a deterministic stub, and the README says so.
~/architecture
- Next.js frontend → FastAPI backend
- Sign in, corpora, upload and question form.
- FastAPI backend → PostgreSQL + pgvector, Redis, MinIO, LLM client
- Ingestion, hybrid retrieval, generation, verification, audit trail.
- PostgreSQL + pgvector
- Chunks, embeddings, corpora, audit records.
- Redis
- Supporting store used by the backend.
- MinIO
- S3-compatible document storage.
- LLM client
- StubLLMClient by default; Anthropic client when a key is configured.
~/stack
- Python 3.12
- FastAPI
- Next.js 16
- PostgreSQL
- pgvector
- Redis
- MinIO
- Docker
~/status
v1.0.0 tagged, AGPL-3.0. Real LLM verification is permanently descoped, so answer quality has not been measured against a real model.
~/related
- lexicon-android Android client for lexicon with offline caching and background sync.