Corpus
1,000+ docs indexed
Live document count from production index
Production citation-backed RAG
Citation-backed RAG that refuses to hallucinate
Ask a real indexed corpus. Every grounded answer shows sources you can open. When retrieval is weak, cite-rag says it does not know.
Next.js 15pgvectorMiniLM 384-dGroqAuth.js
How it works
A fixed pipeline with a grounding gate — not prompt theater.
- 01
Ingest
TXT / PDF corpus via offline CLI
- 02
Chunk
~640 tokens · 12% overlap
- 03
Embed
MiniLM 384-d in-process
- 04
Retrieve
pgvector HNSW · cosine
- 05
Generate
Groq via Vercel AI SDK
- 06
Cite or refuse
Grounded markers — or explicit no
Proof
Live corpus counts and published eval refusal correctness.
Chunks
1003/1003 chunks verified
Scorecard corpus: 1003 documents / 1003 chunks
OOD refusal
100% (8/8)
evals/scorecard.md — out-of-corpus correctly refused
Embeddings
384-dim
Transformers.js all-MiniLM-L6-v2
Architecture
Monorepo seams — UI, RAG library, ingest CLI, vectors, and LLM.
apps/web
Next.js UI · Auth.js · API routes
packages/rag
chunk · retrieve · ground · cite
data/scripts
1,000+ doc ingest CLI
Postgres + pgvector
Supabase · 384-d HNSW
Groq
openai/gpt-oss-20b
Monorepo · designed seams · not a single-file demo
Tech stack
The same stack that runs in production on Vercel + Supabase.
Next.js 15React 19TypeScriptTailwind CSS 4Supabase / pgvectorGroqDrizzle ORMAuth.jsVercel AI SDKTransformers.js