Objective

Parse all 185 existing UDB parameter YAML files and load them into a local searchable database (ChromaDB). This database lets us dynamically find the most relevant existing parameters when building LLM prompts later — instead of hardcoding a few static examples.

Data Sources

• spec/std/isa/param/*.yaml — 207 files total, 185 real parameters (22 are test mocks)
• spec/std/isa/csr/*.yaml — 81 CSR definition files that reference parameters in their IDL code
• spec/schemas/param_schema.json & schema_defs.json — the official schemas

Steps

1. Parse All Parameter YAML Files

Read each parameter file and extract: name, description, long_name, schema (value type), definedBy (which extension), and requirements (IDL constraints). Categorize them by value type:

2. Extract CSR Cross-References

Scan all 81 CSR files and find which CSR fields reference which parameters in their IDL code. For example, mtvec.yaml references MTVEC_MODES inside MODE.sw_write(). Store these cross-references alongside the parameter data — this helps the search engine link related concepts.

3. Build the Vector Database

Convert each parameter into a plain-English summary (combining its name, description, type, extension, and CSR references). Load these into ChromaDB using all-MiniLM-L6-v2 sentence embeddings. Store extension info as filterable metadata.

4. Test Retrieval Accuracy

Write test queries to verify the database returns relevant results:

• "legal values of mtvec mode field" → should return MTVEC_MODES
• "misaligned memory access support" → should return MISALIGNED_LDST
• "width of trap value register" → should return MTVAL_WIDTH

Deliverables

Acceptance Criteria

• All 185 non-mock parameters parsed without errors
• CSR cross-references extracted for parameters used in IDL code
• Test queries return relevant results with >90% accuracy
• Runs entirely offline — no cloud API calls needed