DeepSeekRAG is a retrieval-augmented generation (RAG) application that integrates hybrid retrieval(semantic and keyword search) with the HuggingFace transformers for accurate document question-answering. The application will be tested on the popular DeepSeek-V3 technical report.
- Provide a high-level overview of how the system processes uploads and queries.
- Focuses on key stages: file ingestion, storage, retrieval, generation, and caching.
- Breaks down components and technologies used at each stage.
- Shows specific libraries (PyPDF2, BeautifulSoup, Whoosh, ChromaDB, Hugging Face).
- Includes error handling, ranking, and streaming logic.
-
Web Framework: FastAPI
- Why? it's lightweight, asynchronous (i.e. streaming), and has excellent support for RESTful APIs.
-
Vector Database: Chroma
- Why? It's simple, open-source, and optimized for storing embeddings with semantic search.
-
Embedding Model: Sentence Transformers (e.g., all-MiniLM-L6-v2)
- Why? It’s fast, produces high-quality embeddings for semantic search, and balances accuracy with resource usage.
-
Keyword Search: Whoosh
- Why? it's lightweight and effective for keyword-based retrieval.
-
Language Model: Hugging Face Transformers (e.g.mistralai/Mixtral-8x7B-Instruct-v0.1 via Hugging Face API) . HF Token Needed!
- Why? it's open-source with high-quality generation.
-
PDF/HTML Parsing: PyPDF2 (for PDFs) and BeautifulSoup (for HTML)
- Why? it's reliable for text extraction.
-
Streaming: FastAPI’s StreamingResponse
- Why? It has built-in support for asynchronous streaming.