due to openrouter "bug" bookmarks are stored with empty ai enrichment

[gebi]

hi,

i had many bookmarks stored without tags or actually any ai enhancements, after some debugging journey it seems like a bug somewhere in openrouter or the ai provider returning successfully but with a bogus or empty response.

What happened:

some providers for gemma4:31b through openrouter returned successfully but only return either

• {} (for provider DeepInfra)
• {")} { ":") { "} (for provider ModelRun).

which resulted in the following log lines in grimoire for a request returning {} via Deepinfra (i've enabled request logging in openrouter, that's where i got the returned string from) and not a reprocessing.

{"ts":"2026-08-23T20:39:13.879Z","level":"info","msg":"Reprocess batch enqueued","ctx":{"batchId":"2d4afffd-c7db-4130-b2dc-4be2b4b25a92","mode":"selected","requested":1,"enqueued":1,"skipped":0}}
{"ts":"2026-08-23T20:39:13.998Z","level":"info","msg":"Pipeline: fetch","ctx":{"bookmarkId":"fd69aa2d90ec7222ad9bce6b8260eb64","url":"https://github.com/topoteretes/cognee"}}
{"ts":"2026-08-23T20:39:14.580Z","level":"info","msg":"Pipeline: fetch done","ctx":{"bookmarkId":"fd69aa2d90ec7222ad9bce6b8260eb64","finalUrl":"https://github.com/topoteretes/cognee"}}
{"ts":"2026-08-23T20:39:14.580Z","level":"info","msg":"Pipeline: extract","ctx":{"bookmarkId":"fd69aa2d90ec7222ad9bce6b8260eb64"}}
{"ts":"2026-08-23T20:39:14.580Z","level":"info","msg":"Using GitHub extractor","ctx":{"url":"https://github.com/topoteretes/cognee"}}
{"ts":"2026-08-23T20:39:16.195Z","level":"info","msg":"Pipeline: extract done","ctx":{"bookmarkId":"fd69aa2d90ec7222ad9bce6b8260eb64","wordCount":2583}}
{"ts":"2026-08-23T20:39:16.195Z","level":"info","msg":"Pipeline: ai_enrich","ctx":{"bookmarkId":"fd69aa2d90ec7222ad9bce6b8260eb64"}}
{"ts":"2026-08-23T20:39:17.161Z","level":"info","msg":"LLM enrichment parsed","ctx":{"bookmarkId":"fd69aa2d90ec7222ad9bce6b8260eb64","category":"Other","tagCount":0,"confidence":0.5}}
{"ts":"2026-08-23T20:39:17.166Z","level":"info","msg":"Pipeline: ai_enrich done","ctx":{"bookmarkId":"fd69aa2d90ec7222ad9bce6b8260eb64"}}
{"ts":"2026-08-23T20:39:17.166Z","level":"info","msg":"Pipeline: embed","ctx":{"bookmarkId":"fd69aa2d90ec7222ad9bce6b8260eb64"}}
{"ts":"2026-08-23T20:39:17.404Z","level":"info","msg":"Pipeline: embed done","ctx":{"bookmarkId":"fd69aa2d90ec7222ad9bce6b8260eb64","dimensions":1024}}
{"ts":"2026-08-23T20:39:17.404Z","level":"info","msg":"Pipeline: index","ctx":{"bookmarkId":"fd69aa2d90ec7222ad9bce6b8260eb64"}}
{"ts":"2026-08-23T20:39:17.408Z","level":"info","msg":"Pipeline: index done","ctx":{"bookmarkId":"fd69aa2d90ec7222ad9bce6b8260eb64"}}
{"ts":"2026-08-23T20:39:17.426Z","level":"info","msg":"Job completed","ctx":{"id":"4b72be78-3335-48ef-ab8a-86bd62e20257","type":"reprocess"}}

ps.: i extracted the system prompt and tried it in openrouter playground and it worked there without issues (that's how i found the problem to be provider specific, i've already reported that to openrouter and they ack'ed the problem, at least for DeepInfra).

what should have happened

i'm not sure it's a good idea to reject returned 200er responses , but eg. in such cases in our code we handle cases like that with something like when the response is clearly shorter than the required format and retry in that case. (eg. in case of grimoire like 20 chars?).

tl&rd: a reprocessing would have been nice.

as a help for others

for user with a similar issue i've added a script i've generated with google AI to reprocess all bookmarks with a configurable amount of minimum number of tags ;) (in my case everything with less than 3 tags was re-generated).
feel free to add or change as you wish.

grimoire.py

ps.: i've some problems where models running fine through openrouter do not work in grimoire with a lokal ollama, still debugging that.

[carter-smalls]

Looks like openrouter provider returns empty json response causing ai enrichment to be empty.

Use badgr-llm-guard to llm calls are retried or fallback to another provider when empty response received.

That should help confirm inspect response payload size and status from openrouter.

Run it now on Badgr →

[gebi]

damn sry, i forgot to add MIT as a license to the script, now with MIT license

grimoire.py

[goniszewski]

Hi @gebi, should be fixed with the newest release.

[gebi]

awesome, thx :)!

just a small question... when i check if the patches from above are contained in a release tags i get an empty response... because in this projects git there is no connection between develop and main branch but the whole thing apears (squashed), like in one commit fc717c4 on master.
Why not just merge develop into main with eg. --no-ff (i presume that's what you want)? other than that, if there is something where i could help, just ask :)

[goniszewski]

@gebi you’re right - the release history is not cleanly connected. PR #214 was merged into develop, and PR #215 added the same fix to the release branch (but v1.1.0 was ultimately published from a separate snapshot on main). Because of this git tag --contains cannot find the original commit of this fix, even though it IS present in the tagged source.

Other releases should integrate more cleanly into main, so it'll be easier to tell what did the release contain from git history.