Felt the need to rewrite the entire Godot LLM/RAG/SWE Agent architecture
So far, partial work and deliverables have appeared first, causing the overall baseline to wobble
Decided to first reorganize the whole architecture as a single unit, then design each step’s detailed architecture separately
Plan to separate the flow that runs from data collection, official documentation analysis, chunking, RAG, labeling, SFT/DPO, to the SWE Agent into stages and re‑define it
Organized the Godot official documentation RAG classifier schema and architecture into the main roadmap document
Defined the criteria for converting the Markdown source in outputs/godot_docs_full/pages to JSONL
Documented draft PostgreSQL tables docs_chunks, api_mapping, label_prototypes and their indexing strategy
Recorded the separation of responsibilities and injection order for the AST Parser, Retriever, Validator, and Qwen 3.6
Recorded a retrospective of the questions asked to Qwen about the direction for implementing a Markdown → JSONL converter
Summarized that beyond a simple CLI converter, web upload, API‑key entry, conversion diff checking, schema validation, and multi‑table routing are needed
Stated that a single Markdown file can be mapped simultaneously to docs_chunks, api_mapping, and label_prototypes