sqlite-rag

Minimal RAG: Ollama embeddings + cosine similarity in SQLite. No vector database. ~150 lines, standard library only.

Most RAG stacks pull in a vector database (Chroma, Pinecone, Qdrant, FAISS), a web framework, and a dozen dependencies. For small-to-medium corpora you don't need any of that. This is the whole thing: embeddings come from a local Ollama server, vectors and text live in one SQLite file, and search is plain cosine similarity in Python.

Why

• Zero dependencies — pure Python standard library (sqlite3, urllib, json, math).
• One file, readable in one sitting — easy to audit, fork, and embed in another project.
• Local & private — nothing leaves your machine.
• Portable index — the whole knowledge base is a single .db file you can copy around.

Quickstart

ollama pull nomic-embed-text        # embedding model
python example.py

Example run (real output)

Query: How far from Earth does the telescope orbit, and at which point?
  [0.807] jwst: The James Webb Space Telescope launched on 25 December 2021 ...

Generated answer:
  The telescope orbits the Sun at the second Lagrange point (L2),
  about 1.5 million kilometres from Earth.

Usage

from sqlite_rag import RAG

rag = RAG("mydocs.db")
rag.ingest_file("notes.md")                 # or rag.ingest("some text", source="x")
for hit in rag.search("how do I reset it?", k=4):
    print(hit["score"], hit["source"], hit["text"][:80])

How it works

1. Ingest — text is split into overlapping chunks; each chunk is embedded via Ollama's /api/embeddings and stored in SQLite together with its vector (as JSON).
2. Search — the query is embedded and compared against every stored vector with cosine similarity, computed in Python; the top-k chunks are returned.

That's it. No index structures, no ANN — a linear scan is plenty for thousands of chunks and keeps the code trivial to understand.

nomic-embed-text detail: the model expects task prefixes. This library adds search_document: to stored chunks and search_query: to queries automatically — forgetting these is a common cause of poor recall.

Configuration (env vars)

Variable	Default
OLLAMA_URL	http://127.0.0.1:11434
EMBED_MODEL	nomic-embed-text
CHAT_MODEL (example only)	qwen2.5:7b

Limits (honest)

• Linear scan: fine up to ~tens of thousands of chunks, then add an ANN index.
• Naive fixed-size chunking — swap in your own splitter if structure matters.
• No re-ranking; add a cross-encoder if you need higher precision.

License

MIT

---

See also

Part of a small collection of local-first AI and ESP32 / maker tools:

• sqlite-memory — persistent long-term memory for local LLMs
• ollama-doctor — find out why Ollama is slow (CPU offload / VRAM)
• local-voice-edge — ESP32 voice assistant + local STT→LLM→TTS server
• axs15231b-landscape-lvgl — 3.5" AXS15231B QSPI panel in landscape with LVGL
• guition-esp32p4-lvgl9 — Guition 7" ESP32-P4 + LVGL 9 baseline
• orcaslicer-cli-cookbook — OrcaSlicer from the command line + fixes

⭐ If this saved you time, a star helps others find it.