Review retrieved document excerpts before sending them to a model
At a glance
RAG-Guard's Show HN project demonstrates document Q&A with a manual approval step for retrieved text. This workflow is listed under Document Q&A, Local AI, Human Review.
- Agents / tools
- RAG-Guard, Docker, Ollama, Transformers.js
SourceBuilding a tool for these workflows? See weekly rates.
What they did
1) Have Docker Desktop and a running Ollama endpoint with llama3.2 available. Clone https://github.com/mrorigo/rag-guard.git, enter rag-guard and run ./build.sh. 2) Start the image with port 8000 mapped. For a local Ollama setup, the linked README uses: docker run -p 8000:8000 -e OPENAI_BASE_URL=http://host.docker.internal:11434/v1 -e OPENAI_MODEL=llama3.2 rag-guard. The Ollama endpoint must be reachable from the container. 3) Open http://localhost:8000 and add a .txt or .md document. The browser parses, embeds and searches its text. 4) Ask a document question and inspect the retrieved excerpts. Select and approve only the chunks you want passed to the configured model. 5) Send the approved context and read the returned answer, comparing it with those excerpts.
Worth knowing
The author calls it experimental. The linked implementation supports TXT and Markdown; PDF and DOCX support appear on its future-work list.