System
Architecture
Production-minded RAG for knowledge and support — pipeline exposed, not hidden.
┌───────────────┐
│ Documents │
└───────┬───────┘
↓
┌───────────────┐
│ Parse │
└───────┬───────┘
↓
┌───────────────┐
│ Chunk │
└───────┬───────┘
↓
┌───────────────┐
│ Embeddings │
└───────┬───────┘
↓
┌─────────────────────────────────┐
│ Hybrid Retrieval │
│ Dense ─┐ │
│ ├─ RRF → Rerank │
│ BM25 ──┘ │
└────────────────┬────────────────┘
↓
┌──────────────┐
│ Generate │
└──────┬───────┘
↓
┌───────────────────┐
│ Ground / Refuse │
└─────────┬───────────┘
↓
Answer + CitationsPipeline stages
- 01
Ingest
Fetch docs from catalogued public sources.
- 02
Parse
Structure-preserving HTML, MDX, and Markdown parsers.
- 03
Chunk
Structural chunks tuned for technical documentation.
- 04
Embed
BGE-small dense vectors with optional pgvector store.
- 05
Understand
Route, rewrite, optional HyDE for dense query.
- 06
Retrieve
Dense cosine + BM25 → RRF → feature rerank.
- 07
Generate
Extractive, Ollama, or Bedrock — all grounded.
- 08
Ground
Cite sources or refuse when evidence is insufficient.
Retrieval
Dense + BM25 hybrid with reciprocal rank fusion and lexical reranking.
Generation
Extractive by default (free). Ollama or Bedrock optional behind the same interface.
Storage
In-memory embedding cache at deploy. Optional Postgres + pgvector for scale.
Deployment
Vercel (UI) + Railway (API). Docker Compose for local full stack.