Quick Start Guide

Use this guide to run Big RAG with a small local test collection.

Prerequisites

• LM Studio installed.
• Node.js 18+ and npm.
• An embedding model downloaded in LM Studio, for example nomic-ai/nomic-embed-text-v1.5-GGUF.
• A chat model loaded in LM Studio.

Install and Compile

cd big-rag-plugin
npm install
npx tsc

For full LM Studio development mode, run:

npm run dev

If npm run build fails because .lmstudio/entry.ts is missing, open/run the plugin through LM Studio tooling so that file is generated. npx tsc is enough for local TypeScript validation.

Configure in LM Studio

Required:

• Documents Directory: a directory containing files to index.
• Vector Store Directory: an empty or reusable directory for the Vectra shards.

Recommended defaults for a first run:

Prepare Test Documents

Set Documents Directory to ~/test-docs and Vector Store Directory to a new directory such as ~/.lmstudio/big-rag-test-db.

First Query

Start a chat and ask:

On the first run, Big RAG performs sanity checks, initializes the vector store, scans the directory, indexes supported files, embeds chunks in batches, searches for relevant content, and injects retrieved passages into the prompt.

Subsequent chats reuse the existing index unless the vector store is empty or you request a manual reindex.

Force a Reindex

Turn Manual Reindex Trigger on and send any chat message.

• With Skip Previously Indexed Files on, only new or changed files are indexed.
• With Skip Previously Indexed Files off, all files are rebuilt.

Turn Manual Reindex Trigger off after the run to avoid reindexing on every chat.

Tune After the First Run

• Use sentence chunking for prose-heavy PDFs, DOCX files, and long-form text when retrieval quality matters more than indexing speed.
• Increase Chunk Size for dense technical documentation.
• Raise Min Chunk Length if boilerplate, page numbers, or headers appear in retrieval results.
• Lower Retrieval Affinity Threshold if expected content is not retrieved.
• Reduce Max Concurrent Files or disable OCR if indexing is slow or memory-heavy.

Supported Test Inputs

The current version indexes text-like formats such as .txt, .md, .csv, .json, .yaml, .rst, .log, plus .pdf, .epub, .docx, .odt, HTML, and OCR-enabled images.

Getting Help

• See README.md for full configuration details.
• See EXAMPLES.md for scenario-based setups.
• See TESTING.md for validation steps and performance checks.