fix(rank): keep the official site on page one for known-item searches#92
Merged
Merged
Conversation
ErikChevalier
force-pushed
the
fix/navigational-result-ranking
branch
2 times, most recently
from
438363f to
0705225
Compare
Two re-rank layers could bury a navigational result (the official site of the searched name) below encyclopedia/forum pages even though the aggregator ranked it first. Both are fixed on-device with no new requests. 1. Freshness sort flattened relevance into list position. The default fresh+relevant sort discarded each result's real score (which carries the lexical blend and the navigational boost) and re-derived a flat 1/(60+index) from rank position, so a single dated result could leapfrog an undated official homepage. The aggregator now carries its final score on each result and the freshness blend multiplies THAT, so recency only reorders results of comparable relevance and never overtakes a strong match. 2. AI-slop blocklist false positives. The bundled blocklist is merged from community "hide AI" browser lists that include the official sites of AI companies and major developer hubs (github.com, huggingface.co, openai.com, and more). With the filter on by default those sites were downranked off the first screen. A curated allowlist (resources/blocklist/allowlist.txt) is now subtracted from the effective blocklist at load and build time, including subdomains, so the filter can only sink genuine low-quality domains. Regression tests added for both. Full suite, ruff, and mypy green. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
ErikChevalier
force-pushed
the
fix/navigational-result-ranking
branch
from
0705225 to
f9bbb2e
Compare
Merged
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Problem
Searching a brand or product name returned the official site buried below encyclopedia/forum pages, or off the first screen entirely, even though the metasearch aggregator ranked it first. Diagnosed by running real queries through the live pipeline: this is a precision@1 placement problem in the re-rank layers, not a recall problem (the official site was retrieved and ranked #1, then two later layers demoted it).
Two root causes, both fixed
1. The freshness sort flattened relevance into list position. The default fresh+relevant sort discarded each result's real score (which carries the lexical blend and the navigational boost) and re-derived a flat
1/(60+index)from rank position. That compressed a large lead to a hair, so a single dated result could leapfrog an undated official homepage. The aggregator now carries its final score on each result, and the freshness blend multiplies that, so recency only reorders results of comparable relevance and can never overtake a strong match. (The module docstring claimed it scaled "the relevance score" — it was scaling position.)2. The AI-slop blocklist had high-value false positives, on by default. The bundled blocklist is bulk-merged from community browser "hide AI" lists whose purpose is to keep AI tools out of search results. They include the official sites of AI companies and major developer hubs —
github.com,huggingface.co,openai.com,claude.ai, and 13+ more confirmed. With the filter defaulting todownrank, those sites were pushed off page one. A curated allowlist (resources/blocklist/allowlist.txt) is now subtracted from the effective blocklist at both load and build time (subdomains too), so the filter can only sink genuine low-quality domains.Verification
/api/searchendpoint with the default sort + slop filter on: the official site returns to chore(deps): Bump the github-actions group with 2 updates #1.Notes
relevancescore is kept out of the public JSON contract.urlmrgraded navigational boost, BM25F title>snippet field weighting, registrable-domain dedup to stop one host crowding page one, trust-weighted RRF) are queued separately.🤖 Generated with Claude Code