Update wiki for current v1 retrieval data
4459ac2
docs(wiki): reranking is the largest gain, not a redundancy — and a settings guide
Two pages told readers to skip the single highest-value setting.
Installation-and-Setup called the rerank extra "redundant on an easy corpus
with a strong embedder"; Evidence-Map listed "reranking, candidate pool and
chunk size are NOT the lever". Both came from an early ablation on a small
internal corpus.
PR #103 measured the opposite on the standard public benchmark, with the
DEFAULT local embedder: the largest single retrieval gain anywhere in this
project, intervals disjoint from the baseline. The wiki was advising against
it. Corrected on both pages, with a note saying why the old framing existed
rather than deleting it — it was published advice and someone may have acted
on it. The FINDINGS §7 null is kept and scoped to the corpus it was measured
on, because it is still true there.
New page, Embedders-and-Rerankers: the choice these two settings actually
present. Five embedders (Retrieval-Pipeline said four — OpenAI-compatible was
missing) and two rerankers, each with when to pick it, what it costs, what it
breaks, and a link to the measurement.
It also documents the interaction nothing else does: a candidate pool no wider
than k hands the cross-encoder exactly the list it is meant to reorder. The
flag is set, the model loads, the latency is paid, and the result is identical
to having no reranker — with nothing reporting it.
Canonical-source rule held: the new page states no figure and no default. A
digit scan finds only section references and a numbered list. Verified across
the wiki: 76 repo URLs all resolve, zero broken wikilinks.
5c72fb0
docs(wiki): filtered-ANN tuning is a truncation fix, not a recall win
Both pages described the HNSW filtered-path tuning as fixing "poor recall".
PR #57 measured it twice: truncation is eliminated in both, recall improves
only on a fixture corpus the test rebuilds until it reproduces the pathology,
and moves the other way on a normally-built one. relaxed_order fills to k
with approximate matches — the trade is truncation for approximation.
Same correction landing in README / FINDINGS / CHANGELOG.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
3b51c87
docs(wiki): 14 pages — reference, architecture and evidence routing
The repo already carries the narrative (README, docs/, results/). What it
lacked was a reference surface: nothing documented the 28 RECALL_* variables
or the 8 CLI subcommands in one place, and nothing mapped the modules.
Organised by topic with audience-specific entry paths on Home, so the trust
layer is written once rather than three times.
No page states a measured figure, a config default or a CLI flag value.
Those live in the repo, versioned with the code that produced them, and a
second copy in a wiki — no CI, no review, no commit link — is somewhere for
them to go stale. That is the failure mode this project exists to prevent,
so shipping it in the project's own docs would be self-refuting. Pages
route to FINDINGS / .env.example / --help instead.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
a703da2