2026-06-17 Godot RAG Detector -> Qwen 3.6 Coding Model Development Flow Memo

Structure Image

Core Flow

Godot official documentation preparation  
-> First chunking  
-> Godot-specific post-processing  
-> Build local search infrastructure  
-> Collect and structure GitHub data  
-> Run RAG discriminator  
-> Create training dataset  
-> Train Qwen 3.6 coding model

This note organizes the MVP flow of first creating a Godot RAG detector and then using that detector to generate a training dataset for a Qwen 3.6‑based Godot coding model.

Step Summary

1. Prepare Official Documentation

• Prepare godot_docs_full.zip
• Use the completed crawl of the official documentation
• Currently the data is in its original form, not yet RAG‑processed
• Input consists of a collection of .md files

2. First‑Pass Chunking

• Run chunk_docs.py
• Chunk based on headings
• Large blocks are re‑split using max_chars and overlap
• The first‑pass output is docs_chunks.jsonl

3. Godot‑Specific Post‑Processing

• Remove Sphinx leftovers
• Extract symbols
• Add class/method/property metadata
• Extract migration rules

Outputs:

docs_chunks.jsonl
api_mapping.jsonl
label_prototypes.jsonl

4. Building a Local Search Infrastructure

• Vector DB: document embedding
• Keyword Index: precise search
• Reranker: reordering search results
• API Mapping DB / Label Prototype DB

Important points:

The label is determined by the system.

5. GitHub Data Collection and Structuring

• Collect .gd, .tscn, .tres, project.godot, README
• Build repo tree
• Structure code/scene/configuration files

Deliverables:

GitHub Structured Data JSONL

6. RAG Discriminator Execution

The RAG discriminator uses both the local system and a remote LLM.

Local system responsibilities:

• GitHub chunk input
• symbol extraction
• vector + keyword search
• rerank
• label determination

Remote LLM responsibilities:

• generate corrected code
• generate explanations
• generate QA samples
• generate DPO candidates

The final JSON is assembled by Python code.

7. Generated Dataset

version_classification
api_mapping
migration_fix
instruction_sft
dpo_preference
repo_explorer
patch_generation
metadata_verification

The initial MVP was concluded to be sufficient even if we first generate 10,000–40,000 samples.

8. Model Training

First training:

Qwen 3.6 SFT

Goal:

• Prioritize Godot 4
• Basic GDScript output
• Reject Godot 3 API

Secondary learning:

DPO

Goal:

• Strengthen preference for Godot 4 answers

Future expansion:

SWE Extension

Goal:

• repo explorer
• patch
• trajectory

Core Principles

• Create the RAG discriminator first based on the official documentation.
• The label is determined by the system, not the LLM.
• The LLM is only responsible for generation assistance.
• The ultimate goal is the Qwen 3.6 Godot coding model.