Find the failure before fixing the answer.
The bullets above are reported in the full source resume ↗. Repository findings and any differences are identified separately below.
Ingestion preserves each Reddit comment as a source-labeled chunk and splits long comments into overlapping sentence-aware passages. ChromaDB attaches the same MiniLM embedding function to both indexing and querying, using cosine search over a local persistent collection. Generation discards distant chunks, refuses without an LLM call when none remain, and builds the source list from metadata rather than model-written citations.
FROM THE DEPENDENCY FILESAlternative considered: Blind fixed-size windows or merging unrelated comments
A comment is the document’s natural unit of meaning; preserving it avoids splitting one opinion or diluting several opinions into a single embedding.
Source ↗Alternative considered: Larger hosted or multilingual embedding models
Short English comments suit inexpensive local embeddings; the documented trade-off is weaker recall on domain slang, long passages and other languages.
Source ↗Alternative considered: Tighten the threshold to roughly 0.5
The documented choice avoids masking the known retrieval failure, while the empty-context gate still prevents an unnecessary generation call.
Source ↗Largest chunk 864 → 247 words. The resulting corpus contains 325 chunks across 15 files. This measures chunk size, not answer accuracy.
Source ↗The Stat 134 question refused despite relevant material being in the corpus: top-four retrieval favored unrelated courses and a contentless comment. Human retrieval-quality and answer-accuracy columns are still unfilled, so no scored before/after gain is claimed.
Source ↗Excerpts from the repository’s own notes. These limits belong beside the results.
No tracked Python test_ functions found; evaluation scripts and notebooks are described separately. Source ↗
Inspect comment delimiters, long-comment splitting and source metadata.
Follow the distance gate and programmatic source attribution.
Read the five fixed questions and the human-review boundary.