You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
research(atlas): boundary-drift measurement — how much a paragraph's vector/rank moves with different embedding neighbors #6038
Span reuse assumes a span's stored vector stays valid when the span is reused in a new version. But an embedding is a function of the chunk's exact text including its window neighbors: the same paragraph embedded with different surrounding text is a different vector, and its rank for a given query can move. How much?
Scope
Take a sample of real corpus paragraphs; embed each in its original window and in perturbed windows (shifted boundaries, changed neighbors); measure cosine drift + rank movement on a fixed query set.
Deliverable is a number with a distribution: median/p95 drift, and the reuse-safety rule it implies (e.g. reuse only byte-identical spans whose recorded context matches; re-embed when context class changes).
Done looks like
A short measurement report + the reuse rule it justifies, linked from the span-schema design (sibling under #6028).
Verdict: valid, distinct research gap; keep #6038 open as the scoped research follow-up. The request is not implemented or superseded. Fresh checks: #6038 OPEN; no hard-stop labels; no linked PR (the only search hit, #4708, is unrelated and merged). Parent #6028 research notes list #6038 as a parallel measurement gate; the qualitative discussion does not provide the requested measurements.
Current-main evidence: laptop main was fetched and fast-forwarded (already up to date); local HEAD and GitHub main both equal 50237a7debf6e9114122c593a15a467858d2bf9a. pkgs/atlas-search/crates/atlas-embed/src/lib.rs:149-170 embeds input text; crates/atlas-embed/tests/gate_a.rs:57-139 tests fixed probes/model parity, not changed-neighbor drift. crates/atlas-core/eval.rs:118-125 consumes precomputed document rankings; eval/README.md:61-70 says its smoke fixture is synthetic and no real held-out set is in the harness. No boundary-drift report or experiment was found.
Scoped route: run the bounded offline experiment in #6038: sample real paragraph-bearing chunks; keep model and query set fixed; compare each target embedded in original vs perturbed windows; report cosine drift and vector-rank/top-K movement (median/p95, split by perturbation or source type), then state a reuse rule and link results from #6035. Keep corpus text private; synthetic fixtures are not evidence. This can proceed in parallel with #6035 schema design, whose byte-identical reuse/context-reference contract should consume the result. No code changes or duplicate queue packet: #6038 itself is the existing scoped follow-up issue. No tests/builds run.
Activity
developing-today commentedon Nov 5, 2026
CODE-1703
dezren39 commentedon Nov 5, 2026
🤖 Agent note · handed off
35 remaining items
dezren39 commentedon Nov 5, 2026
🤖 Agent note · handed off
dezren39 commentedon Nov 5, 2026
🤖 Agent note · picked up
dezren39 commentedon Nov 5, 2026
Triage — task 10404 (2026-11-05)
Verdict: valid, distinct research gap; keep #6038 open as the scoped research follow-up. The request is not implemented or superseded. Fresh checks: #6038 OPEN; no hard-stop labels; no linked PR (the only search hit, #4708, is unrelated and merged). Parent #6028 research notes list #6038 as a parallel measurement gate; the qualitative discussion does not provide the requested measurements.
Current-main evidence: laptop main was fetched and fast-forwarded (already up to date); local HEAD and GitHub main both equal
50237a7debf6e9114122c593a15a467858d2bf9a.pkgs/atlas-search/crates/atlas-embed/src/lib.rs:149-170embeds input text;crates/atlas-embed/tests/gate_a.rs:57-139tests fixed probes/model parity, not changed-neighbor drift.crates/atlas-core/eval.rs:118-125consumes precomputed document rankings;eval/README.md:61-70says its smoke fixture is synthetic and no real held-out set is in the harness. No boundary-drift report or experiment was found.Scoped route: run the bounded offline experiment in #6038: sample real paragraph-bearing chunks; keep model and query set fixed; compare each target embedded in original vs perturbed windows; report cosine drift and vector-rank/top-K movement (median/p95, split by perturbation or source type), then state a reuse rule and link results from #6035. Keep corpus text private; synthetic fixtures are not evidence. This can proceed in parallel with #6035 schema design, whose byte-identical reuse/context-reference contract should consume the result. No code changes or duplicate queue packet: #6038 itself is the existing scoped follow-up issue. No tests/builds run.