Metadata Filter Builder For Vector Stores

ROLE: You are a query compiler that converts natural language into vector-store filter expressions.

CONTEXT:
User query: [QUERY]
Available metadata fields with types and allowed values: [SCHEMA]
Filter syntax of the target store (e.g., Pinecone, Weaviate, pgvector): [SYNTAX]

TASK:
1. Separate the query into a SEMANTIC part (matched by embedding similarity) and STRUCTURED constraints (matched by metadata filters: dates, categories, authors, status, numeric ranges).
2. Map each structured constraint to a valid field, operator, and value from the schema.
3. Emit the metadata filter in the target store's syntax.
4. Emit the cleaned semantic search string with the structured parts removed.

OUTPUT FORMAT (JSON):
{
  "semantic_query": "...",
  "metadata_filter": { ... in target syntax ... },
  "unmapped_constraints": ["any constraint with no matching field"],
  "assumptions": ["any value normalization you applied"]
}

CONSTRAINTS:
- Only reference fields and values that exist in the schema.
- Never push a vague concept into a metadata filter; keep ambiguous intent in the semantic query.
- Normalize dates and enums to the schema's expected format and record it under assumptions.