Skip to content

Resolve paper titles that differ from the catalog by a typo - #29

Merged
NielsRogge merged 1 commit into
mainfrom
claude/fuzzy-paper-title-lookup
Oct 15, 2026
Merged

NielsRogge merged 1 commit into
mainfrom
claude/fuzzy-paper-title-lookup

Conversation

@NielsRogge

Copy link
Copy Markdown
Collaborator

Why

Production chat audit (2026-10-15): get_paper_info("BEIR: A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models") returned not_found. The catalog holds the arXiv spelling "Heterogenous" (2104.08663), which keyword search returns first, but title resolution only accepts exact titles or a short name.

What

  • near_title(): for a full title (≥ 24 normalized characters), a search result counts as a near match when the normalized titles have a SequenceMatcher ratio ≥ 0.95 and the same numbers (so "Llama 3" ≠ "Llama 4", and other editions or years don't match).
  • _resolve_paper falls back to a unique near match after exact and short-name matches. Several near matches → the existing "ambiguous" error. URL slugs never use it.
  • pwc-mcp 0.2.8.

Validation

  • uv run pytest -q tests in mcp_server: 71 passed, including new BEIR, ambiguous-twin, and short-name/edition rejection tests.
  • Replayed last week's 144 distinct failed production title lookups against the live catalog (phrase query plus the first page of the plain query): 2 now resolve correctly (BEIR; "DINO: Emerging Properties in Self-Supervised Vision Transformers" → 2104.14294). There were no false matches; the rest are papers outside the catalog.

Chat-related: stops after review. It ships via pwc-deploy <env> <sha> --component mcp only with explicit authorization.

🤖 Generated with Claude Code

Production chat asked get_paper_info for "BEIR: A Heterogeneous Benchmark
..." and got not_found, although the catalog holds the arXiv spelling
"Heterogenous" (2104.08663) and keyword search returns it first.

A full title (24+ normalized characters) now also resolves to a unique
search result whose normalized title matches at a 0.95 similarity ratio
with the same numbers, so typos and variant spellings resolve while
sequels, other editions, and short names do not. Several near matches
stay ambiguous. Replaying last week's 144 failed title lookups against
the live catalog resolves BEIR and "DINO: Emerging Properties in
Self-Supervised Vision Transformers" with no false matches; the rest
are papers outside the catalog.

Bumps pwc-mcp to 0.2.8.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@NielsRogge
NielsRogge merged commit f69ef14 into main Oct 15, 2026
2 checks passed
@NielsRogge
NielsRogge deleted the claude/fuzzy-paper-title-lookup branch October 15, 2026 07:48
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant