01
Corpus + chunking audit
We characterise your corpus before we pick a chunker. PDFs with tables, code with comments, transcripts with timestamps — each gets a different strategy.
★ RAG architectureDirect LLM · no framework8-week engagement
RAG · RETRIEVAL-AUGMENTED GENERATIONHybrid retrieval. Citations, not hallucinations.
Postgres + pgvector for vector search, BM25 for lexical, fused with reciprocal-rank fusion. Citation-first answers — every claim links to its chunk. No separate vector DB, no five-system synchronisation problem.
PostgreSQLpgvectorTypeScriptOpenAI
Cycle
8 weeks · corpus to citation
Stack
Postgres · pgvector · TypeScript
Output
Ingestion + retrieval + citation surface
Discipline
Citations on every answer, no exceptions
[WHY THIS EXISTS]
Most RAG demos fall apart at scale.
A naive cosine-similarity search over a 1M-chunk corpus retrieves plausible-but-wrong context, the LLM hallucinates confidently on top, and the user discovers the failure mode by getting fired for citing it. Hybrid retrieval + citations make the failure mode visible.
- Pgvector + BM25 hybrid — dense semantic + lexical exact-match
- Reciprocal-rank fusion to merge scores defensibly
- Citation surface — every claim in the answer links to its source chunk
- Chunking pipeline tuned to your corpus, not someone's blog post
[HOW WE BUILD IT]
Retrieval first, generation second.
02
Postgres + pgvector
One database for your text, your embeddings, your metadata, your access controls. Skip the vector-DB-shaped tax.
03
Hybrid retrieval
BM25 finds the exact-term matches the embeddings miss; vector search finds the semantic matches BM25 misses. Reciprocal-rank fusion merges them.
04
Citation-first answers
The prompt requires the model to cite each claim. The UI renders inline citations. Users (and auditors) can verify every assertion.
[OUTCOMES AT HANDOFF]
What's live at week eight.
100%
Answers with inline citations
1 DB
Postgres for text + vectors + metadata
Hybrid
BM25 + pgvector with RRF fusion
Per-tenant
Access control on retrieval, not just generation
[ALSO WORTH READING]
Related LLM engagements.
DIRECT LLM · APPLIED KBring the problem.
Bring the problem.
We’ll bring the build.
Eight weeks, fixed price, eval suite at handoff. Direct LLM engineering on top of the K-Framework. Two Q3 slots remain.