本地知识库语义搜索引擎一键设计提示词

You are a senior AI infrastructure engineer. Design a fully local semantic search engine for my document collection.

My Setup:

• OS: [Mac/Linux/Windows]
• GPU: [NVIDIA RTX xxxx / Apple Silicon M-series / CPU only]
• Documents: [describe your docs - PDFs, markdown, code, meeting notes, etc.]
• Scale: [approximate number of documents and total size]

Design Requirements:

1. Document Ingestion Pipeline

• List exact tools for each document format (PDF to text, DOCX, HTML, code files)
• Chunking strategy: recommend chunk size, overlap, and method (semantic vs fixed)
• Metadata extraction: titles, dates, authors, tags

2. Embedding and Vector Store

• Recommend the best local embedding model for my hardware
• Compare: Nomic Embed, BGE, GTE, E5 - pick one with justification
• Vector DB: recommend between ChromaDB, Qdrant, LanceDB for local use
• Index configuration: HNSW parameters, quantization settings

3. Query Pipeline

• Implement hybrid search (vector + BM25 keyword)
• Re-ranking with a local cross-encoder model
• Query expansion/reformulation strategy

4. QA Layer

• Local LLM selection for answer generation (Qwen, Llama, Phi)
• Context window management: how to fit retrieved chunks
• Citation/source attribution in responses

5. Implementation Plan

Provide a step-by-step bash/python script outline I can run to set this up. Include all pip install commands, model download commands, and config files.

Output: Complete technical design document with copy-paste ready commands.