Commit graph

869 commits

Author SHA1 Message Date
Yiorgis Gozadinos
6c5bc0aae1
Gate combined-chat capability vision on the driving model 2026-07-24 15:27:20 +03:00
Yiorgis Gozadinos
597808c56e
Fix chat analysis-model selection, AG-UI example state, and chat docs
Drive analysis-only chat with analysis.model (falling back to qa.model) so
the running model matches the one the analysis capability configures,
including its vision flag; RAG-bearing chats still run on qa.model.

Give the AG-UI example state-bearing deps and a final STATE_SNAPSHOT so
registered citations reach the client, mirroring the app backend.

Update chat docs: haiku-rag chat uses --capability/-c, and drop the removed
"View state" command-palette entry.
2026-07-24 15:27:20 +03:00
Yiorgis Gozadinos
b8dcb066dc
Port cite partial-success feedback to capabilities 2026-07-24 15:27:19 +03:00
Yiorgis Gozadinos
175929f23a
Improve coverage 2026-07-24 15:26:18 +03:00
Yiorgis Gozadinos
77585ef26a
fix per-question capability limits and tool isolation 2026-07-24 15:26:18 +03:00
Yiorgis Gozadinos
43c17a6777
fix capability execution limits and chat loading 2026-07-24 15:26:17 +03:00
Yiorgis Gozadinos
22ffd5e92e
Add max retries=3 back 2026-07-24 15:26:17 +03:00
Yiorgis Gozadinos
9deb1f2bd4
replace haiku.skills with native Pydantic AI capabilities 2026-07-24 15:26:17 +03:00
Yiorgis Gozadinos
35b8b046ce
Cover the detached-chunk guard in the multimodal rerank search test 2026-07-24 13:07:39 +03:00
Yiorgis Gozadinos
543aba7547
Multimodal reranking: send picture chunks to vllm rerankers as images
reranking.multimodal (vllm provider only) attaches picture bytes to
synthetic picture chunks before rerank; VLLMReranker sends them as
content-parts documents (base64 data URI + description text) in the
same /v1/rerank request as plain text documents.
2026-07-24 12:29:20 +03:00
Yiorgis Gozadinos
0308220769
Read verbs skip the embeddings config-compatibility check 2026-07-24 12:14:31 +03:00
Yiorgis Gozadinos
d9cc9a4eea
Narrow AutoTokenizer type for transformers 5 2026-07-23 12:01:28 +03:00
Yiorgis Gozadinos
75a21c82c2
Search results and citations carry document metadata 2026-07-23 10:23:48 +03:00
Yiorgis Gozadinos
8bf569ef89
cite reports unresolvable chunk ids on partial success 2026-07-22 11:56:56 +03:00
Yiorgis Gozadinos
3034280c78
Skip embedding validation on document deletion 2026-07-22 10:48:00 +03:00
Yiorgis Gozadinos
1ed424b50d
Serialize metadata-only document updates with the write lock 2026-07-16 13:11:35 +03:00
Yiorgis Gozadinos
f02cde5ddf
Fix same-tick cancellation losing recovery results in _wait_protected 2026-07-16 13:11:35 +03:00
Yiorgis Gozadinos
dd2817ff6d
Protect tag-creation cleanup from cancellation 2026-07-16 13:11:35 +03:00
Yiorgis Gozadinos
1f88944ada
Protect restore rollback from cancellation; report delete_tag listing failures 2026-07-16 13:11:34 +03:00
Yiorgis Gozadinos
ce1b9ac88e
Harden history against tag annotation failures; document tag restore 2026-07-16 13:11:34 +03:00
Yiorgis Gozadinos
5453c00c95
Test restore and migration separation 2026-07-16 13:11:34 +03:00
Yiorgis Gozadinos
4522fbdf1b
Add tag restore CLI 2026-07-16 13:11:34 +03:00
Yiorgis Gozadinos
f41ec379df
Add Store.restore_tag 2026-07-16 13:11:34 +03:00
Yiorgis Gozadinos
a1f3435df3
Vacuum suppresses OSError only 2026-07-16 13:11:34 +03:00
Yiorgis Gozadinos
f997d7adc9
Consolidate tag primitives 2026-07-16 13:11:34 +03:00
Yiorgis Gozadinos
0e271eaf4b
Remove --before/--at time travel 2026-07-16 13:11:33 +03:00
Yiorgis Gozadinos
0813a1c980
Add tag CLI commands and history tag annotations 2026-07-16 13:11:27 +03:00
Yiorgis Gozadinos
0cbde6b7a2
Serialize tag operations, vacuum, and metadata refresh with writes 2026-07-16 12:32:57 +03:00
Yiorgis Gozadinos
69319c9390
Grow vacuum retention to protect tagged versions 2026-07-16 12:32:57 +03:00
Yiorgis Gozadinos
ae603d9b4d
Add database-level tag primitives to Store 2026-07-16 12:32:56 +03:00
Yiorgis Gozadinos
aa620aeb60
Bump lancedb to 0.34.0 2026-07-16 12:32:56 +03:00
Yiorgis Gozadinos
d2f5ecfc58
Prevent magic-byte sniffing from misrouting text conversion in docling-local
Docling guesses a DocumentStream's format from its first bytes before
considering the extension, so markdown/HTML content starting with a binary
magic signature ("BM" = BMP, "ID3" = MP3) was routed to an image or audio
backend and fell back to plain-text conversion. Prefix the encoded text
with a newline so the sniff finds nothing and the extension decides.
2026-07-15 19:20:54 +03:00
Yiorgis Gozadinos
144900d385
Remove the mxbai reranking provider 2026-07-14 11:09:55 +03:00
Yiorgis Gozadinos
5e2a928013
Allow transformers 5.x in the mxbai extra 2026-07-14 11:09:55 +03:00
Yiorgis Gozadinos
82db15cf3c
Fill missing item-table pages from fully-visible inputs in context expansion 2026-07-14 10:32:45 +03:00
Yiorgis Gozadinos
ee5b7b7313
Split merged context-expansion groups that cannot afford every constituent 2026-07-14 10:15:28 +03:00
Yiorgis Gozadinos
271fdf9b5a
Add Logfire debugging skills and worker-breaker event 2026-07-10 13:23:17 +03:00
Yiorgis Gozadinos
a82673a900
Route eval scripts through the shared telemetry configure 2026-07-10 11:38:33 +03:00
Yiorgis Gozadinos
73ba764922
Emit a docling_serve.request span per instance attempt 2026-07-10 11:16:27 +03:00
Yiorgis Gozadinos
da79f92af1
Honor OTEL_SERVICE_NAME and set service name/version for all processes 2026-07-10 11:11:46 +03:00
Yiorgis Gozadinos
18c0f6c5e8
Add --no-expand to the visualize command for chunk-only grounding 2026-07-09 12:49:42 +03:00
Yiorgis Gozadinos
995d081aa1
Visualize the exact context the model saw via Citation.doc_item_refs
visualize_chunk re-expanded chunks from scratch to recover their refs,
which could not faithfully reproduce the original merge, scores, and
clip — so a visualization could highlight different pages than the
citation covered. Carry the cited items on Citation.doc_item_refs and
resolve bounding boxes from them directly; re-expansion remains only as
the fallback for callers with no stored context (CLI, inspector). Chat,
inspector, the app endpoint, and the frontend pass the refs through.
2026-07-09 11:51:37 +03:00
Yiorgis Gozadinos
c424f51056
Anchor and clip merged citations on the highest-scoring chunk
A merged search result took its chunk_id from whichever constituent
sorted earliest in the document, while its score was the max across the
group — so the citation's identity could point at a different, less
relevant chunk. Anchor chunk_id and the content/refs fallbacks on the
max-score constituent, clip the budget window around that same chunk so
its evidence is never trimmed away, and narrow page_numbers, doc_item_refs,
and attached image bytes to the items that survive the clip.
2026-07-09 11:05:41 +03:00
Yiorgis Gozadinos
6d86237dd6
Draw matched content stronger than expanded context in visualizations 2026-07-09 11:05:41 +03:00
Yiorgis Gozadinos
494f774473
Visualize all constituent chunks of a merged citation 2026-07-09 11:05:41 +03:00
Yiorgis Gozadinos
0bcf34363a
Carry merged chunk ids on SearchResult and Citation 2026-07-09 11:05:41 +03:00
Yiorgis Gozadinos
b14152a45b
Keep picture and table hits section-bounded during context expansion 2026-07-09 11:05:40 +03:00
Yiorgis Gozadinos
c02bcd5dc9
Pool HTTP clients for vLLM embedding and vLLM/Jina reranking 2026-07-09 10:32:44 +03:00
Yiorgis Gozadinos
43a580afb3
Dedupe picture chunks and skip small pictures at chunking 2026-07-09 10:24:32 +03:00
Yiorgis Gozadinos
a1cf405cae
Rename document_meta identity column document_id to id 2026-07-08 16:32:55 +03:00