mini-RAG

Ask questions about your documents with mini-RAG. Built in about 20 lines of shell on top of this open source stack: Stack: Docling (parsing) → TF-IDF (retrieval) → Ollama (local LLM)

docs/ ──docling──▶ tmp/md/*.md ──TF-IDF top-k──▶ tmp/ctx ──Ollama──▶ answer

Requirements

• macOS or Linux with Homebrew
• Python 3
• jq: brew install jq

Quick start

1. Start the LLM server (terminal 1). Skip this if the Ollama app is already running.

sh server.sh

2. Ask a question (terminal 2)

sh rag.sh ./docs "What are geospatial foundation models?"

Usage

sh rag.sh <docs_path> "<question>" [top_k] [model] [ollama_url]

Arguments are positional (there are no --flags), so they must be given in this order:

Position	Argument	Description	Default
1	docs_path	File or folder of documents	required
2	question	Your question, in quotes	required
3	top_k	Number of chunks sent to the LLM	4
4	model	Ollama model name	$MODEL or qwen2.5:7b
5	ollama_url	Ollama server address	$OLLAMA_URL or http://localhost:11434

Example with 8 chunks:

sh rag.sh ./docs "What are geospatial foundation models?" 8

Environment variables (optional):

MODEL=llama3.1:8b sh rag.sh ./docs "Question?"      # use another model
OLLAMA_URL=http://gpu-box:11434 sh rag.sh ...       # remote Ollama server

Debugging

Intermediate files are saved in ./tmp/:

File	Content
tmp/md/	Markdown produced by Docling, one file per document
tmp/ctx	The retrieved chunks sent to the LLM (last run only)

If an answer looks wrong, run cat tmp/ctx to check whether the right chunks were retrieved. To force a clean re-parse, run rm -rf tmp/md.

⚠️ The first launch is slow. Dependencies are installed, Docling downloads its models, and Ollama pulls the LLM (~4.7 GB for qwen2.5:7b).