idea_world_labDEV JOURNAL
Monday, June 22, 2026

June 22, 2026

  • Added a local PostgreSQL setup for the Godot official‑document RAG classifier
    • Documented how to run a pgvector/pgvector:pg16 container using Docker Compose.
    • Aligned the DB payload columns with the field names of docs_chunks.jsonl, api_mapping.jsonl, and label_prototypes.jsonl.
    • Defined tables docs_chunks, api_mapping, label_prototypes, ingest_reports and keyword/exact‑search indexes.
    • Kept the embedding column open; the actual vector index will be created in a separate migration after the embedding model dimensions are finalized.
    • DB setup documentation: docs/database/2026-06-22-local-postgres-setup.md
  • Recorded a retrospective of the Markdown → JSONL conversion and local DB ingestion preparation process
    • Described the workflow of converting official‑document Markdown to intermediate JSONL for preview/verification before loading into the DB.
    • Explained why the conversion results are split into docs_chunks, api_mapping, and label_prototypes and structured for easy insertion into local PostgreSQL.
    • Retrospective: docs/retrospectives/2026-06-22.md
  • Documented the reason for making the repository public and then private again, and reflected on the decision to publish
    • Although I felt insecure about my current skill level and found publishing burdensome, I concluded that sharing the record can help others and also accelerate my own growth.
    • Also outlined a plan to run a local LLM endpoint on an Oracle Cloud 24 GB VRAM instance and attach it to GitHub workflow/PR review automation.
    • Because I lost access to the Oracle Cloud account and found RunPod setup cumbersome, I postponed LLM‑based PR review automation for later.
  • Decided to focus today on converting roughly 1,500 official‑document Markdown files to JSONL and loading them into the local DB
    • Converting a single file takes longer than expected, so it is still uncertain whether all 1,500 can be finished today.
    • Based on current logs, 39 files are done and 4 are deferred (total 43 files) after about 1 hour 9 minutes of processing.
    • The average speed is about 1.6 minutes per file, so converting all 1,570 files is estimated to require roughly 42 hours of compute time.
  • Since tomorrow and the day after are likely unavailable, I outlined the verification plan for the subsequent work
    • Plan to inject the already generated JSONL into local PostgreSQL and verify that it is searchable.
    • According to the workflow in docs/roadmaps/2026-06-21-initial-rag-classifier-architecture.md, a small‑scale test will be run from a Python script: source code → AST Parser → Retriever → evidence JSONL → Qwen 3.6 API call.
    • Periodically check how Qwen 3.6 responds based on the Retriever evidence.