mpep-search（mpep 搜索）

MPEP Search Skill

Search MPEP corpus through hybrid RAG (FAISS vector + BM25 keyword + HyDE + cross-encoder reranking).

Sources:

• MPEP: Manual of Patent Examining Procedure
• 35 USC: United States Code Title 35
• 37 CFR: Code of Federal Regulations Title 37
• Subsequent Publications: Federal Register updates (post-Jan 2024)

Core Operations

1. search_mpep

Inputs:

• query (string, required): Search query (minimum 3 characters)
• top_k (int, optional): Number of results (default: 5, max: 20)
• retrieve_k (int | None, optional): Candidates before reranking (default: top_k * 4, max: 100)
• source_filter (string | None, optional): Filter by source ("MPEP", "35_USC", "37_CFR", "SUBSEQUENT", or None)
• is_statute (bool | None, optional): Filter for statute content
• is_regulation (bool | None, optional): Filter for regulation content
• is_update (bool | None, optional): Filter for recent updates

Outputs:

{
    "rank": int,
    "source": str,
    "section": str,
    "file": str,
    "page": int,
    "has_statute": bool,
    "has_mpep_ref": bool,
    "has_rule_ref": bool,
    "is_statute": bool,
    "is_regulation": bool,
    "is_update": bool,
    "relevance_score": float,
    "text": str,
    # Optional for SUBSEQUENT:
    "doc_type": str,
    "fr_citation": str,
    "effective_date": str
}

Examples:

# Basic search
search_mpep("enablement requirement 35 USC 112", top_k=5)

# Search only statutes
search_mpep("written description", top_k=10, is_statute=True)

# Search recent updates
search_mpep("AI inventorship", is_update=True)

# Filter by source
search_mpep("fee schedule", source_filter="37_CFR")

2. get_mpep_section

Retrieve all content from specific MPEP section.

Inputs:

• section_number (string, required): MPEP section number (e.g., "2100", "608.01")
• max_chunks (int, optional): Maximum chunks to return (default: 50)

Outputs:

{
    "section": str,
    "total_chunks": int,
    "chunks": [
        {
            "text": str,
            "metadata": {
                "source": str,
                "file": str,
                "page": int,
                "section": str,
                "has_statute": bool,
                "has_mpep_ref": bool,
                "has_rule_ref": bool,
                "is_statute": bool,
                "is_regulation": bool,
                "is_update": bool
            }
        }
    ]
}

Error Response:

{"error": "No content found for MPEP section {section_number}"}

Examples:

# Get MPEP 2100 (Patentability)
get_mpep_section("2100", max_chunks=50)

# Get subsection
get_mpep_section("608.01")

Input Validation

Query validation:

• Minimum 3 characters
• Case-insensitive
• No empty/whitespace-only queries

Section number validation:

• Numeric with optional decimal (e.g., "100", "2100", "608.01")

Limits:

• top_k capped at 20
• retrieve_k capped at 100

Implementation Notes

Index Location:

• FAISS index: mcp_server/index/mpep_index.faiss
• Metadata: mcp_server/index/mpep_metadata.json
• BM25 index: mcp_server/index/mpep_bm25.json

Search Architecture:

1. HyDE Query Expansion (hypothetical documents)
2. Hybrid Retrieval (FAISS vector + BM25 keyword via RRF)
3. Cross-Encoder Reranking (final relevance scores)
4. Metadata Filtering (source/type filters)

Dependencies:

• sentence-transformers (BGE-base-en-v1.5)
• FAISS (vector search)
• rank-bm25 (keyword search)
• Cross-encoder (reranking)
• HyDE (optional, graceful degradation)

Error Handling:

• Clear error messages for missing index/invalid queries
• Graceful degradation if HyDE fails
• Input validation before processing