Similarity vs. relevance
What was the amount I was pre-approved for when I got my mortgage from Wells Fargo?
VECTOR SEARCHcos 0.6830
“pre-approved for $350,000 from Wells Fargo.”
similarity to the current claim
SUPERSEDED FACT
answer_3a6f1e82_1 · 2023-08-11 · retrieved FIRST for this question at cos 0.7457 over 304 session chunks
ERRATA · BELIEF GRAPHas of the fixture
$425,000
the claim no SUPERSEDES points away from · user-correction
no session chunk contains it at all — it was appended as a correction, so retrieval over the transcript cannot return it at any cosine · embedder: BAAI/bge-small-en-v1.5
the claim no SUPERSEDES points away from · user-correction
no session chunk contains it at all — it was appended as a correction, so retrieval over the transcript cannot return it at any cosine · embedder: BAAI/bge-small-en-v1.5
“Similar, sure. Relevant? Almost never.”
The vector pane reads apps/web/fixtures/beat-0.94.json — measured on 2026-08-18 by eval/embed_beat.py against BAAI/bge-small-en-v1.5 over history 852ce960-clean, committed, so the demo makes no network call to an embedding service. The 0.6830 is the cosine between the two claims’ own evidence spans; 0.7457 is what the same embedder retrieves first for the question above, and it is the retired claim’s session. The belief pane is not a fixture: it is /api/belief at page load, the same read the pre-approval answer and its revision chain are built from.