← All selected work
GraphRAG · Retrieval engineering
GraphAtlas
A knowledge platform that fuses keyword, dense-vector, and graph retrieval, then makes quality and provenance measurable.
ScopePortfolio build using a synthetic organizational corpus and a 50-question evaluation set.
- hybrid retrieval
- 0.840 Hit@5
- hybrid latency
- 537 ms p95
- faithfulness
- 4.38 / 5
- PostgreSQL 17
- pgvector
- Hono
- TypeScript
- React
- LightRAG
01 / Context
The problem
A RAG demo can produce an answer without proving where it came from or whether retrieval improved. GraphAtlas needed citations, graph context, and an evaluation loop that could compare retrieval strategies.
02 / Architecture
How the system fits together
- 01DocumentsPDF, Word, XML, MD, CSV
- 02Parsernormalized source spans
- 03LightRAGentity + relation extraction
- 04PostgreSQLpgvector + graph schema
- 05Keyword / Vector / Graphparallel recall
- 06RRFrank fusion + router
- 07Agent APISSE + citations
Supporting systems
- React graph explorerone-hop expansion
- Benchmark harness50-question golden set
03 / Engineering
What I built
- Built local ingestion for PDF, Word, XML, JSON, Markdown, and CSV files, preserving section-level source spans and rejecting failed parses instead of indexing partial documents.
- Used a LightRAG-based Python extraction sidecar to create the knowledge graph and load it into PostgreSQL.
- Combined PostgreSQL full-text search, pgvector HNSW search, and cycle-safe graph traversal with Reciprocal Rank Fusion.
- Served retrieval and streaming answers through a Hono/TypeScript API with a hand-written tool-calling agent and chunk-level citations.
- Built a React graph explorer, benchmark dashboard, versioned golden set, four-mode ablation, and traceable result files.
04 / Evidence
Measured outcomes
hybrid Hit@5
Improved from 0.800 with vector-only retrieval on the 50-question set.
GraphAtlas README — measured results ↗faithfulness
LLM-judged answer faithfulness for hybrid retrieval.
GraphAtlas benchmark ↗p95 latency
Measured for hybrid retrieval on the same golden set.
GraphAtlas benchmark ↗05 / Demo
Product evidence
06 / Boundaries
Trade-offs and limitations
- The corpus and organizational entities are synthetic.
- The evaluation set is small, uses an LLM judge, and does not establish production traffic performance or multi-tenant security.
07 / Sources