The initially written 27th document contained many object and structure descriptions, making it difficult to trace how the input actually flows
Then I readjusted the document direction based on PR feedback
Modified it to explicitly state, for each expanded input line # <relative path>, which file and line go into the AST Parser
Using actual line ranges such as E020‑E034 in player.gd, I organized the flow where that portion of the text leads to Retriever search and LLM judgment
File paths are for tracing only, and I fixed the rule that only the chunkText of code/text fragments goes into Retriever search
docs_chunks, api_mapping, label_prototypes are searched in the same way without special treatment, and the search candidates are re‑validated by the LLM
Today’s work aimed to lock down, in documentation, how input and output connect at each point before actual implementation, so that the AI does not arbitrarily change ranges or flow into abstract structural explanations
As of now, chunk‑level decomposition is considered reasonably successful; the next key focus is how to actually perform DB searches
Need to verify which JSONL candidates return from docs_chunks, api_mapping, label_prototypes using only chunkText
Must determine how Qwen verification decides relevance and discards candidates based on the retrieved JSONL
Before adding DB search, I used GPT to create demo Godot chunks and related/unrelated JSONL to experiment with matching prompts
Initially I asked “Does this JSONL contain content that corresponds to the source code? Answer only yes/no,” but both related and unrelated JSONL were answered “yes”
Then I limited “yes” to cases where at least one of source_api, source_pattern, match_terms, required_when_seen_in_code, before_code directly matches the actual SOURCE_CODE string/API call
By avoiding broad word similarity or the LLM’s prior Godot knowledge and looking only at the string evidence in the JSONL, related JSONL returned “yes” and unrelated returned “no”
This experiment showed that the Qwen verification step after DB search should first judge whether the retrieved JSONL contains a string directly matching the current chunk, rather than relying on “plausible semantic similarity”
Tomorrow I will create several demo sets of Godot chunks with related/unrelated JSONL and repeatedly test how Qwen derives “yes”/“no” based on evidence
