Also deleted the draft scripts for chunking/post‑processing/validation and the initial RAG chat/index draft
Left work/godot_rag/ empty for now, to avoid using the faulty outputs as a baseline
Recorded the reason for the cleanup as a retrospective
Determined that the problem was proceeding with chunking before sufficiently analyzing the original official‑documentation structure
Noted that the goal of using the entire Godot official documentation was polluted by the LLM’s arbitrarily chosen “MVP core API” and hard‑coded/over‑engineered directions
Stated that, without my instruction, ChatGPT/Codex would arbitrarily narrow the scope or assume an answer existed before a step was finished and produce the next output
Concluded that when LLM‑generated content, regex candidates, official‑documentation evidence, and user‑approved rules are mixed, they cannot be used as a basis for labeler/RAG discriminators
Re‑recorded the current architecture’s shortcomings
Static‑analysis layer is weak
No GDScript AST/parser‑based validation
Godot project dependency graph is weak
Execution/syntax validation is weak
Label taxonomy is still coarse
Distinguishing provenance between LLM‑generated content and verified answers is weak
Data‑leakage/duplicate‑removal design is weak
Reset the next work direction to the original analysis of godot_docs_full
Instead of immediately rebuilding the RAG or catalog, we will first analyze outputs/godot_docs_full/pages itself
Need to verify what Markdown structure each class reference, migration, and tutorial document has
Must redesign which chunk granularity to use—page‑level, section‑level, or API‑member‑level—according to the Godot official‑documentation structure
Decided that a validation‑report standard should be defined before chunking