← portfolio

RAG Playground — self-hosted · zero external services

checking model…

Corpus

＋ Upload

.txt / .md / .pdf · ≤ 5 MB. Files are chunked (~700 chars, 120 overlap), embedded locally (bge-small ONNX), and indexed into SQLite FTS5 + vector store.

How it works

Ingest — text extraction → paragraph-aware chunking → local ONNX embedding → SQLite (FTS5 + float32 vectors)
Retrieve — BM25 keyword leg + cosine vector leg, fused by Reciprocal Rank Fusion (k=60)
Cite — every result links back to its source chunk; "Ask" mode generates strictly from retrieved context

Bring-your-own-key answers: your key is used for that single request only — never logged or stored.

generated · grounded in citations

built from scratch · FastAPI + SQLite FTS5 + fastembed (bge-small-en ONNX) · runs on a 2-core VPS ← back to portfolio