Research overview

This section documents the research phase of knowledge — the exploration that produced the design decisions now used in production. The work was carried out as six Jupyter notebooks (nb01–nb06); each page walks through one notebook: the question it asked, what was tried, and what it concluded.

How the experiments were evaluated

A shared, deterministic method runs across the notebooks so results are comparable:

• Gold query sets — curated question→expected-source pairs, including a 200-question bilingual (FR / EN) set, used to score retrieval objectively.
• Deterministic retrieval metrics — recall@k, source coverage and related measures, shared across nb02–nb06 (not eyeballed).
• Token tracking — bundle/context token counts, to compare strategies on cost as well as accuracy.

The six notebooks

#	Topic	Outcome	Status
nb01	Corpus discovery & filtering	Profile-driven file selection for the corpus	done
nb02	Markdown chunking	Selected: AST-merge + breadcrumb, chunk_size=500	done
nb03	JS & Vue SFC chunking	Structure-aware chunkers + BM25/RRF hybrid retrieval	done
nb04	JSON chunking	Category-aware key splitting	done
nb05	Embedding model evaluation	Recommends Qwen3-Embedding-0.6B	done
nb06	Qdrant index & evaluation	Real Qdrant index, verified against nb05; Qdrant / Chroma / LanceDB comparison	done

How the findings feed production

• Chunking winners (nb02–nb04) → the ingestion pipeline.
• Embedding choice (nb05) → RAG: indexing & retrieval.
• Index & retrieval validation (nb06) → RAG: indexing & retrieval.

Notebooks & code

The notebooks and their helper code live in the repository under docs/experiments/ (notebooks/nb0X_*.ipynb plus the matching *_lab/ helpers). They are not part of the rendered site — these pages are the readable write-up of what they contain.