pending — blocked on Phase 1 close
Phase 2 — DuckDB + VSS + Multi-view Embeddings
Ingest all YAML cards into a single DuckDB file. Compute three embedding views per card using local llama.cpp pipelines, store as HNSW-indexed VSS columns, and validate retrieval quality against 20 hand-authored query pairs before Phase 3 can begin.
← Back to site mapWhat Phase 2 will build
Storage
DuckDB
prime_radiant.duckdb with VSS extension. One row per card. Schema mirrors Pydantic model with JSON columns for nested fields.semantic_emb
768-dim gemma3-embedding over
description. HNSW index. Primary retrieval vector.state_emb
~50-dim deterministic embedding from state variable values + phase one-hot. Enables structural similarity search.
graph_emb
~64-dim node2vec over the full coupling graph. Captures relational neighborhood structure.
Toolchain
| Component | Tool | Config key |
|---|---|---|
| Database | DuckDB + VSS extension | data/prime_radiant.duckdb |
| Embedding model | gemma3-embedding via llama.cpp | LLAMA_EMBEDDING_BIN, EMBEDDING_MODEL_PATH |
| Graph embedding | node2vec (networkx + pecanpy) | Walk length 80, p=1, q=1 |
| Smoke test | scripts/verify_stack.py |
Already validates DuckDB+VSS plumbing |