progressed the official‑doc JSONL conversion collection from 1,550 to 1,422 items using the Markdown → JSONL Converter
continued running Qwen Validation Debugger from item 46 to 50, checking generated code, Godot 3/4 engine inspection, and JSONL validation results
completed common code and JSONL validation for item 50; item 48’s dynamic scene‑resource loading kept failing engine checks because it generated a fixed resource path that does not exist
when retrying, passed the previous engine diagnosis and improved the debugger to use a different stable seed each generation round without manually registering syntax or answers per item
stored the current JSONL of collector 8501 to outputs/godot-rag-sqlite/godot-rag.sqlite3 and cross‑checked the 1,570 original URLs·domains·SHA‑256 and record sources from pages.zip
built docs_chunks, api_mapping, label_prototypes, original provenance, and an FTS5 search index in SQLite, allowing snapshot recreation with the same builder before the full collection finishes
after comparing SQLite and vector‑DB dumps of the collected JSONL for Git upload, chose SQLite that preserves original records and provenance and can be regenerated/shared without a separate server
initially added SQLite to Git only as a baseline artifact, keeping the Source Flow Debugger’s F strategy pointing to PostgreSQL, which created consistency issues requiring separate reflection of the same JSONL in both SQLite and PostgreSQL
to eliminate dual management, merged the F strategy’s baseline data and execution repository into a single committed SQLite, removing PostgreSQL connection settings and migration paths
BM25 candidates are now read from SQLite FTS5, computed as Okapi BM25 in Node, and embeddings are stored together with model·dimension·record‑content SHA‑256 and Float32 vectors in SQLite’s record_embeddings, re‑indexing only changed records
the Source Flow Debugger now directly checks the committed SQLite’s revision, record count, and embedding count without a separate DB URL, and the stored vectors are searched via cosine similarity in Node
vector‑DB dumps depend on the embedding model, dimension, and index implementation, so they are not kept as separate baseline data; when needed they are regenerated as derived indexes from the same SQLite