Removes classify_retrieval_intent and the type/folder filter parameters on
retrieve_context. The keyword classifier was the same anti-pattern as the
formatting-driven docx chunker: a heuristic that locks the user into specific
phrasings and fails silently on anything novel. A scope enum (personal /
library / conversations / memory) would have been the same heuristic in a
fancier wrapper — the categories themselves are mine, not Aaron's.
New shape: a retrieve_documents tool exposed to Claude. Tool takes a single
query argument; the model decides when to call it, what to search for, and
how many times per turn (multi-query falls out naturally for compound asks).
Pre-LLM retrieval is gone — memory still rides as ground truth in the prompt,
but corpus content is fetched on demand by the model with concrete queries
it crafts itself, not the user's raw phrasing.
retrieve_context is now pure: hybrid retrieval + cross-encoder rerank + dedup,
no filters. The reranker ranks, the model judges relevance. When ranking
fails (e.g. abstract instructional queries pulling philosophy books), the
right fix is a better reranker, not another query-time taxonomy. That work
is acknowledged but deferred.
System prompt updated to teach the model about the tool and to prefer
concrete tokens (named entities, project names, course codes) over abstract
phrasing when constructing search queries.
Three refinements to retrieve_context, all keyed off observed failures from
test_retrieval.py:
- Library/personal split. classify_retrieval_intent now returns
(type_filter, folder_exclude_prefixes). Biographical document intent excludes
Library/* so philosophy/cognition books stop crowding out CVs and dossiers
for queries like "write me a bio".
- Near-duplicate collapse. Multi-folder copies of the same file (e.g., several
Teaching Philosophy.pdf in different application folders) used to fill the
top-N with the same content. Dedup by first-300-chars hash after rerank.
- Folder in source citations. Surface metadata.folder alongside basename so
the LLM can disambiguate among 21 CV.docx variants and the user can see
which copy a citation refers to.
Also: bump hnsw.ef_search to 500 when a WHERE filter is present.
pgvector 0.6 doesn't iterate past its initial HNSW candidate list, so a
restrictive filter that excludes the nearest neighbors otherwise returns
empty.
Replaces pure-dense top-8 retrieval with a three-stage pipeline:
- BM25 (tsvector + websearch_to_tsquery) and dense (pgvector) in parallel,
fused with Reciprocal Rank Fusion
- Optional type filter driven by classify_retrieval_intent() so questions
about prior conversations don't pull documents and vice versa
- Cross-encoder rerank (ms-marco-MiniLM-L-6-v2) over RRF candidates before
taking final top-N
Also adds scripts/reindex_docx_pptx.py — one-off re-ingest used to recover
table/header/text-box content in docx and pptx after the 93c0d89 extractor
upgrade — and scripts/test_retrieval.py to exercise the new pipeline against
representative queries.
Schema: requires GIN index on to_tsvector('english', document) (already
created out-of-band via psql since Apache AGE in shared_preload_libraries
blocks ALTER TABLE on this database).
aaron