Spaces:
Running
Running
| title: "distill-rag" | |
| emoji: "🧰" | |
| colorFrom: "indigo" | |
| colorTo: "blue" | |
| sdk: "static" | |
| pinned: false | |
| tags: | |
| - tool | |
| - rag | |
| - dataset | |
| - nlp | |
| - text-processing | |
| - search | |
| - embeddings | |
| # Distill RAG | |
| A lightweight pipeline for extraction, chunking, embeddings, and search. | |
| --- | |
| ## 🧠 What Is This? (Plain-Language Overview) | |
| **distill_rag** is a small but powerful toolkit that helps you transform messy text sources (HTML pages, transcripts, articles, archives) into **clean, structured data** that AI models can learn from. | |
| It’s designed for people who want to build: | |
| * **domain-specific AI assistants** | |
| * **high-quality expert models** | |
| * **distillation pipelines** where a stronger model teaches a smaller one | |
| If you’ve ever tried to fine-tune a model and realised the hardest part is actually preparing the dataset — this toolkit is for that problem. | |
| --- | |
| ## 🌍 Why This Exists | |
| Training or distilling a specialised AI model requires **clean, coherent, well-structured data**. | |
| But most text online is: | |
| * full of ads, scripts, headers | |
| * chopped into small fragments | |
| * badly formatted | |
| * missing metadata | |
| * hard to chunk in meaningful ways | |
| Before you can train a model, you need a **pipeline that turns raw text into polished, training-ready data**. | |
| `distill_rag` gives you that pipeline. It helps you: | |
| 1. pull content from raw HTML | |
| 2. clean and structure it | |
| 3. break it into long coherent chunks | |
| 4. embed it locally | |
| 5. index it with Elasticsearch | |
| 6. perform high-quality semantic search | |
| This structure mirrors the data format most distillation workflows expect. | |
| **Goal:** | |
| Make it easy for researchers and builders to create **high-quality domain-specific AI models**. | |
| --- | |
| ### ⚡ Key Benefits: Speed, Simplicity, and Efficiency | |
| Built entirely in Node.js, `distill_rag` leverages async promises and lightweight concurrency to deliver blazing-fast performance—often 5–10× faster than equivalent Python-based tools like LlamaIndex, LangChain, or Haystack. This makes it ideal for local workflows on consumer hardware, where you can process thousands of chunks in minutes without heavy dependencies or complex setups. | |
| - **GPU-Bound Efficiency**: Embeddings (via Ollama) are the bottleneck, but everything else (extraction, chunking, indexing) is lightning-fast, with sustained rates of 30+ chunks/second on a single RTX 3090. | |
| - **No Overhead Bloat**: Direct HTTP calls to embedding APIs and seamless parallelism mean low CPU/RAM usage and no GIL-like bottlenecks. | |
| - **Hackable for JS Devs**: If you're in the Node ecosystem (e.g., integrating with web apps or CLIs), this fits like a glove—zero language switching required. | |
| - **Benchmark Example**: On an RTX 3090 with 1531 sessions (~3565 chunks, 5000–9000 chars each), a full index rebuild takes just 1m48s (33 chunks/s). Comparable Python tools often take 8–40 minutes for similar workloads due to wrapper latencies and inefficient batching. | |
| If speed and developer joy matter in your RAG/distillation pipeline, this toolkit shines. | |
| --- | |
| ## 🔧 How It Works (At a Glance) | |
| ### 1. Extract & Clean | |
| Feed it a folder of HTML (scraped, archived, downloaded). | |
| It removes noise and produces structured JSON sessions. | |
| ### 2. Chunk & Embed | |
| Text is broken into long, context-rich chunks (ideal for distillation). | |
| Each chunk is embedded using a local model like `mxbai-embed-large`. | |
| ### 3. Index & Search | |
| Chunks are stored in Elasticsearch as vectors + metadata. | |
| You can then run semantic search to retrieve relevant material. | |
| --- | |
| ## 🚀 Who Should Use This? | |
| * AI researchers building **aligned distilled models** | |
| * Developers training **expert assistants** | |
| * Archivists handling large text collections | |
| * Anyone building **custom RAG systems** | |
| * Anyone who wants a clean, open, hackable indexing pipeline | |
| --- | |
| ## 🚀 Features | |
| * ✔ HTML extraction using Cheerio | |
| * ✔ Robust cleaning (scripts, ads, headers removed) | |
| * ✔ Long-chunker tuned for distillation | |
| * ✔ Async embedding + indexing | |
| * ✔ Elasticsearch v8 dense vector support | |
| * ✔ Ollama embedding API support | |
| * ✔ CLI tools for extraction, indexing, search | |
| * ✔ Test suite included | |
| * ✔ Apache 2.0 licensed | |
| --- | |
| ## 🧱 Project Structure | |
| ```text | |
| distill_rag/ | |
| ├── data_extraction/ | |
| │ ├── clean_html.js # strip noise safely | |
| │ ├── extractor.js # extract Q/A style turns | |
| │ ├── convert_raw_to_sessions.js # HTML → structured JSON | |
| │ └── walk_and_extract.js # CLI to batch-convert directories | |
| │ | |
| ├── indexing/ | |
| │ ├── index_distill_chunks.js # long-chunk indexer | |
| │ └── rebuild_distill_index.sh # wipe + rebuild helper | |
| │ | |
| ├── search/ | |
| │ ├── search_distill_chunks.js # BM25 / vector / hybrid search | |
| │ └── search_cli.js # CLI search tool | |
| │ | |
| ├── tests/ # jest-based automated test suite | |
| │ | |
| ├── prompts/ # optional prompt templates | |
| ├── shared/ # shared utilities | |
| ├── cleanup.sh # remove build artefacts | |
| ├── jest.config.js | |
| ├── package.json | |
| └── README.md | |
| ``` | |
| --- | |
| ## 📦 Installation | |
| Requirements: | |
| * Node **18+** | |
| * Elasticsearch **8.x** | |
| * Embedding API (e.g., **Ollama** running `mxbai-embed-large`) | |
| Install: | |
| ```bash | |
| npm install | |
| ``` | |
| --- | |
| ## 🧼 1. Extracting Data from Raw HTML | |
| Convert a directory: | |
| ```bash | |
| node data_extraction/walk_and_extract.js raw_html/ extracted_sessions/ | |
| ``` | |
| Output example: | |
| ```json | |
| { | |
| "title": "example.html", | |
| "turns": [ | |
| { "role": "user", "content": "Q: What is service?" }, | |
| { "role": "assistant", "content": "Service begins with kindness." } | |
| ] | |
| } | |
| ``` | |
| Behind the scenes it: | |
| * strips scripts, ads, headers, nav bars | |
| * extracts paragraphs and headings | |
| * assigns roles (`user` = first block, rest `assistant`) | |
| * normalises whitespace | |
| --- | |
| ## 🧱 2. Chunking + Indexing into Elasticsearch | |
| Rebuild the full index: | |
| ```bash | |
| bash rebuild_distill_index.sh | |
| ``` | |
| Manual index build: | |
| ```bash | |
| ES_DISTILL_INDEX=quo_distill_index \ | |
| JSON_DIR=./extracted_sessions \ | |
| ELASTICSEARCH_NODE=http://localhost:9200 \ | |
| node indexing/index_distill_chunks.js | |
| ``` | |
| The indexer: | |
| * creates long semantic chunks (5000–9000 characters) | |
| * calls your embedding API (`/api/embeddings`) | |
| * indexes all chunks with metadata: | |
| * `title` | |
| * `session_date` | |
| * `source` | |
| * `chunk_index` | |
| * `embedding` (vector) | |
| --- | |
| ## 🔎 Advanced Retrieval Modes | |
| `distill_rag` supports **three complementary search strategies** via `search/search_distill_chunks.js`: | |
| ### 1. BM25 (Keyword Search) | |
| Classic lexical search. | |
| Good for names, citations, exact phrases. | |
| ```js | |
| const { searchBM25 } = require("./search/search_distill_chunks"); | |
| const results = await searchBM25("service to others", 5); | |
| console.log(results); | |
| ``` | |
| --- | |
| ### 2. Vector Search (Dense Embeddings) | |
| Semantic similarity using your local embedding model. | |
| ```js | |
| const { searchVector } = require("./search/search_distill_chunks"); | |
| const results = await searchVector("how to grow spiritually", 5); | |
| console.log(results); | |
| ``` | |
| --- | |
| ### 3. Hybrid Search (RRF Fusion) — Recommended | |
| State-of-the-art fusion of: | |
| * BM25 lexical relevance | |
| * Dense vector KNN | |
| This gives robust results even on noisy or varied datasets. | |
| ```js | |
| const { searchHybrid } = require("./search/search_distill_chunks"); | |
| const results = await searchHybrid("balance love and wisdom", 5); | |
| console.log(results); | |
| ``` | |
| --- | |
| ## 🖥 Search CLI Tool | |
| You can run searches directly from the terminal: | |
| ```bash | |
| node search/search_cli.js "service to others" | |
| ``` | |
| Specify mode (`bm25`, `vector`, `hybrid`) and `k`: | |
| ```bash | |
| node search/search_cli.js "healing catalyst" hybrid 8 | |
| node search/search_cli.js "unity" bm25 5 | |
| node search/search_cli.js "wisdom" vector 10 | |
| ``` | |
| This prints: | |
| * source file | |
| * chunk index | |
| * score | |
| * preview of the retrieved chunk | |
| --- | |
| ## 🧪 Tests | |
| Run all tests: | |
| ```bash | |
| npm test | |
| ``` | |
| Covers: | |
| * HTML cleaning | |
| * extractor correctness | |
| * session conversion | |
| * chunker behaviour | |
| * embedding API live test | |
| * Elasticsearch index live test | |
| * hybrid retrieval | |
| * end-to-end smoke test | |
| --- | |
| ## 🧽 Cleanup | |
| ```bash | |
| npm run clean | |
| ``` | |
| or: | |
| ```bash | |
| bash cleanup.sh | |
| ``` | |
| --- | |
| ## 🛠 Configuration | |
| Config is handled via environment variables: | |
| | Variable | Default | Purpose | | |
| | -------------------- | --------------------------------------- | ------------------------------- | | |
| | `ELASTICSEARCH_NODE` | `http://localhost:9200` | ES cluster URL | | |
| | `ES_DISTILL_INDEX` | `quo_distill_index` | Target index | | |
| | `EMBED_URL` | `http://localhost:11434/api/embeddings` | Embedding API | | |
| | `EMBED_MODEL` | `mxbai-embed-large` | Embedding model | | |
| | `CHUNK_MIN` | `5000` | Minimum chunk size (characters) | | |
| | `CHUNK_MAX` | `9000` | Maximum chunk size (characters) | | |
| | `JSON_DIR` | *(required)* | Directory of session JSON | | |
| --- | |
| ## 📄 License | |
| Apache 2.0 (see `LICENSE`). | |
| --- | |
| ## 🤗 Contributing | |
| Contributions are welcome — bug fixes, new extractors, support for other embedding backends, indexing strategies, documentation. | |
| --- | |
| ## ☀️ Final Thoughts | |
| This project is meant to empower people building truth-aligned, service-oriented models. | |
| If it helps someone create a clearer dataset or a kinder AI, it’s doing its job. | |
| --- | |