okami.mom/web-extract-go

git clone https://git.nkp.pet/okami.mom/web-extract-go

nekomimi.petweb-extract: the Go service running on the Mac mini (vassago)5a90548

main
3.8 KiB33 linesraw

jev batching · 2026-09-21

incident

the five searches ending at 2026-09-20 17:56:32 utc made 1,439 jev calls: 92 reranking calls and 1,347 passage calls. four shared slots and a 60 ms global pacing interval left them running until the 90-second request deadline. recorded jev latency summed to 351 seconds across those four slots. search results were found; no summarizer calls were made. successful page results were then discarded by the transport because the overall operation returned a deadline error.

46ad8cd5ee12bdd28f0d5b87926f4731 alone made 263 calls (20 ranking, 243 passage), not 200: the dashboard shows at most 200 decision rows. its individual calls generally took about a quarter-second, not 90 seconds each.

change

typesafe's api accepts structured instructions and independently evaluated questions in one request. rank candidates and passage judgments now use that mechanism, capped at 96 questions and 100,000 encoded bytes. question-local data prevents unrelated passages from entering another question's context. answer ids are mapped back to original items; oversized batches split without truncating data or dropping checks. eight in-flight batch slots are independent of the four page/summarizer workers. pacing and bounded retries remain.

deadline handling preserves successful sources and fills unfinished per-url statuses. partial failures are visible in request logs rather than being reported as fully successful. historical rows, users, credentials, completion request shapes and prose task names are unchanged.

verification and limits

  • unit/race/vet checks cover batching, byte/count bounds, answer mapping, concurrency/cancellation, raw-input preservation, and partial deadline responses. frontend request and real-browser fixture tests pass.
  • frozen replay of the exact 263 recorded inputs: 17 api calls, 992 decisions, 4.93 seconds for jev work. this excludes searching and fetching. input usage grew from 180,154 to 246,273 tokens because question-local passage data is repeated across four checks; fewer requests do not imply fewer billed tokens.
  • question placement changes scores: that replay retained 13 of 17 formerly eligible passages and selected 23 overall. average absolute probability change was 0.037. historical votes are not ground-truth relevance labels, so these are diagnostics, not proof of identical retrieval quality. the first attempt's arbitrary 90% old-vote agreement check failed and was removed rather than presented as a quality benchmark; required-fact checks remain separate.
  • frozen required-fact tests retained sqlite's single-writer correction and go's %w/Is/As evidence. the first live summary omitted %w despite it being in selected evidence; a repeat passed. this remains a stochastic summarization limitation, not a claim of perfect preservation.
  • one development replay got a non-retried upstream http 400; later replay and the five-request staging run completed with no jev errors. error handling continues to fail closed for invalid or missing judgments.

five concurrent searches used the original queries, 20 requested results, and the original focused-highlight query on a loopback-only staging server on the mini. full original request bodies were not stored, so this is a production-shaped replay, not an exact replay of every client option; live pages/search results can change.

original request prefixsecondsresultsjev calls
46ad8cd5ee129.591519
caa99d9a503c10.182024
41de4f16e34410.432024
307d122efde39.381314
a0c531ebd1ab8.602023

all returned usable results without a request-level timeout or jev error. this does not eliminate slow/blocked source pages or guarantee a latency bound under arbitrary load.

1# jev batching · 2026-09-21
2
3## incident
4
5the five searches ending at 2026-09-20 17:56:32 utc made 1,439 jev calls: 92 reranking calls and 1,347 passage calls. four shared slots and a 60 ms global pacing interval left them running until the 90-second request deadline. recorded jev latency summed to 351 seconds across those four slots. search results were found; no summarizer calls were made. successful page results were then discarded by the transport because the overall operation returned a deadline error.
6
7`46ad8cd5ee12bdd28f0d5b87926f4731` alone made 263 calls (20 ranking, 243 passage), not 200: the dashboard shows at most 200 decision rows. its individual calls generally took about a quarter-second, not 90 seconds each.
8
9## change
10
11[typesafe's api](https://docs.typesafe.ai/api) accepts structured instructions and independently evaluated questions in one request. rank candidates and passage judgments now use that mechanism, capped at 96 questions and 100,000 encoded bytes. question-local data prevents unrelated passages from entering another question's context. answer ids are mapped back to original items; oversized batches split without truncating data or dropping checks. eight in-flight batch slots are independent of the four page/summarizer workers. pacing and bounded retries remain.
12
13deadline handling preserves successful sources and fills unfinished per-url statuses. partial failures are visible in request logs rather than being reported as fully successful. historical rows, users, credentials, completion request shapes and prose task names are unchanged.
14
15## verification and limits
16
17- unit/race/vet checks cover batching, byte/count bounds, answer mapping, concurrency/cancellation, raw-input preservation, and partial deadline responses. frontend request and real-browser fixture tests pass.
18- frozen replay of the exact 263 recorded inputs: **17 api calls, 992 decisions, 4.93 seconds** for jev work. this excludes searching and fetching. input usage grew from 180,154 to 246,273 tokens because question-local passage data is repeated across four checks; fewer requests do not imply fewer billed tokens.
19- question placement changes scores: that replay retained 13 of 17 formerly eligible passages and selected 23 overall. average absolute probability change was 0.037. historical votes are not ground-truth relevance labels, so these are diagnostics, not proof of identical retrieval quality. the first attempt's arbitrary 90% old-vote agreement check failed and was removed rather than presented as a quality benchmark; required-fact checks remain separate.
20- frozen required-fact tests retained sqlite's single-writer correction and go's `%w`/`Is`/`As` evidence. the first live summary omitted `%w` despite it being in selected evidence; a repeat passed. this remains a stochastic summarization limitation, not a claim of perfect preservation.
21- one development replay got a non-retried upstream http 400; later replay and the five-request staging run completed with no jev errors. error handling continues to fail closed for invalid or missing judgments.
22
23five concurrent searches used the original queries, 20 requested results, and the original focused-highlight query on a loopback-only staging server on the mini. full original request bodies were not stored, so this is a production-shaped replay, not an exact replay of every client option; live pages/search results can change.
24
25| original request prefix | seconds | results | jev calls |
26| --- | ---: | ---: | ---: |
27| `46ad8cd5ee12` | 9.59 | 15 | 19 |
28| `caa99d9a503c` | 10.18 | 20 | 24 |
29| `41de4f16e344` | 10.43 | 20 | 24 |
30| `307d122efde3` | 9.38 | 13 | 14 |
31| `a0c531ebd1ab` | 8.60 | 20 | 23 |
32
33all returned usable results without a request-level timeout or jev error. this does not eliminate slow/blocked source pages or guarantee a latency bound under arbitrary load.