RAGLite

RAGLite is a local-first CLI that turns a folder of docs (PDF/HTML/TXT/MD) into structured, low-fluff Markdown — and then makes it searchable with Chroma 🧠 + ripgrep 🔎.

Security note (prompt injection)

RAGLite treats extracted document text as untrusted data. If you distill content from the web (or any third party), assume it may contain prompt injection attempts. The distillation prompts explicitly instruct the model to ignore any instructions inside source material and treat sources as data only.

Core idea: compression-before-embeddings ✂️➡️🧠

What you get

For each input file:

• *.execution-notes.md — practical run/operate notes (checks, failure modes, commands)
• *.tool-summary.md — compact index entry (purpose, capabilities, entrypoints, footguns)

Optionally:

• raglite index stores embeddings in Chroma 🧠 (one DB, many collections)
• raglite query runs hybrid search 🔎 (vector + keyword)

Why local + open-source?

If you want a private, local setup (no managed “fancy vector DB” required), RAGLite keeps everything on your machine:

• Distilled Markdown artifacts are plain files you can audit + version control
• Indexing uses Chroma (open-source, local) and keyword search uses ripgrep
• You can still swap in a hosted vector DB later if you outgrow local

Engines

RAGLite supports two backends:

• OpenClaw (recommended): uses your local OpenClaw Gateway /v1/responses endpoint for higher-quality, format-following condensation.
• Ollama: uses POST /api/generate for fully local inference (often less reliable at strict templates).

Prereqs

• Python 3.11+
• An LLM engine:
  • OpenClaw (recommended) 🦞, or
  • Ollama 🦙
• For search:
  • Chroma (open-source, local) 🧠 at http://127.0.0.1:8100

Install

# from repo root
python3 -m venv .venv
source .venv/bin/activate
pip install -e .

Quickstart (60s)

# 0) Setup
cd ~/Projects/raglite
source .venv/bin/activate

# 1) Condense → Index (one command)
raglite run /path/to/docs \
  --out ./raglite_out \
  --engine ollama --ollama-model llama3.2:3b \
  --collection my-docs \
  --chroma-url http://127.0.0.1:8100 \
  --skip-indexed

# 2) Query
raglite query ./raglite_out \
  --collection my-docs \
  "rollback procedure"

Usage

1) Distill docs ✍️

raglite condense /path/to/docs \
  --out ./raglite_out \
  --engine openclaw

(Or fully local: --engine ollama --ollama-model llama3.2:3b.)

2) Index distilled output (Chroma)

raglite index ./raglite_out \
  --collection my-docs \
  --chroma-url http://127.0.0.1:8100

3) Query (hybrid)

raglite query ./raglite_out \
  --collection my-docs \
  --top-k 5 \
  --keyword-top-k 5 \
  "rollback procedure"

Useful flags

• --skip-existing : don’t redo files that already have both outputs
• --skip-indexed : don’t re-embed chunks that are already indexed
• --nodes : write per-section nodes + per-doc/root indices
• --node-max-chars 1200 : keep nodes embed-friendly
• --sleep-ms 200 : throttle between files (helps avoid timeouts)
• --max-chars 180000 : cap extracted text per file before summarizing

Output layout

RAGLite preserves folder structure under your --out dir:

<out>/
  some/subdir/file.execution-notes.md
  some/subdir/file.tool-summary.md

(Default output folder is `./raglite_out`.)

Notes / gotchas

• PDF extraction is best-effort: scanned PDFs without embedded text won’t be great.
• If you use --engine openclaw, pass --gateway-token or set OPENCLAW_GATEWAY_TOKEN.
• Indexing defaults to high-signal artifacts (nodes/summaries/notes) and skips *.outline.md unless you opt in.

Roadmap

Current (implemented)

• condense — condense/summarize documents into Markdown artifacts
• index — chunk + embed + store in Chroma collections
• query — retrieve relevant chunks (vector + keyword)
• run — one-command pipeline (condense → index)
• Outline + nodes + indices: --outline, --nodes, root index.md + per-doc *.index.md

Next (near-term)

• Detect deletions (prune removed chunks from Chroma)
• Batch upserts to Chroma for speed
• Better query output formatting (snippets + anchors)
• raglite doctor (dependency checks)

(Full: ROADMAP.md)

---

Built to turn “docs” into usable, searchable tool knowledge.