Commit graph

951 commits

Author SHA1 Message Date
Yiorgis Gozadinos
f4ddb0f5c6
Allow chat agent to specify the search limit 2026-01-12 15:52:40 +02:00
Yiorgis Gozadinos
2da2c6d13a
Search tool in chat agent needs no search context 2026-01-12 15:13:43 +02:00
Yiorgis Gozadinos
5adab04e50
Update prompts 2026-01-12 15:10:04 +02:00
Yiorgis Gozadinos
d43821659c
Filter low-confidence answers 2026-01-12 15:01:32 +02:00
Yiorgis Gozadinos
e4a7d86348
Rank q/a history with respect to cosine similarity to new queries. Adapt the chat agent to only pass relevant q/as 2026-01-12 14:15:36 +02:00
Yiorgis Gozadinos
a9ed178a90
Research graph test cassette 2026-01-12 12:43:51 +02:00
Yiorgis Gozadinos
7cc561d1db
Refactor to make agents a top-level module. Bring in the conversational agent from the app 2026-01-12 12:37:08 +02:00
Yiorgis Gozadinos
fc29b0c3f7
Update docs 2026-01-12 12:37:08 +02:00
Yiorgis Gozadinos
bb61e8fab6
Remove custom AG-UI infrastructure in favor of pydantic-ai native support 2026-01-12 12:37:07 +02:00
Yiorgis Gozadinos
baefbe9416
Tool calling in the frontend 2026-01-12 12:36:32 +02:00
Yiorgis Gozadinos
f2fb64cbca
Remove custom ag-ui from app backend. 2026-01-12 12:36:32 +02:00
Yiorgis Gozadinos
26bc02ef14
Build a "conversational" research graph 2026-01-12 12:36:31 +02:00
Yiorgis Gozadinos
df71e7a893
Adapt search to show complete results 2026-01-12 12:36:31 +02:00
Yiorgis Gozadinos
8579ded7bb
Natural language filtering 2026-01-12 12:36:31 +02:00
Yiorgis Gozadinos
36f964eaa9
get_document() 2026-01-12 12:36:31 +02:00
Yiorgis Gozadinos
a6a60963c0
Add SSE heartbeat to prevent connection timeouts
Sends SSE comment every 15 seconds during long LLM operations
to keep the connection alive and prevent body timeout errors.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
2026-01-12 12:36:30 +02:00
Yiorgis Gozadinos
8b7c679159
Add SearchAgent for internal query expansion
SearchAgent generates 2-4 diverse search queries internally,
runs them against the knowledge base, deduplicates by chunk_id,
and returns consolidated results. This prevents the outer chat
agent from making multiple search calls.

Also updates system prompt to enforce single tool calls per message.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
2026-01-12 12:36:30 +02:00
Yiorgis Gozadinos
49f5c20757
Mount haiku.rag.yaml config in docker compose
🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
2026-01-12 12:36:30 +02:00
Yiorgis Gozadinos
18cec25fc9
Visual grounding; 2026-01-12 12:36:30 +02:00
Yiorgis Gozadinos
72b66dd3d8
Styling 2026-01-12 12:36:30 +02:00
Yiorgis Gozadinos
6ba28244d8
Logfire integration for debugging 2026-01-12 12:36:29 +02:00
Yiorgis Gozadinos
14873862f8
Q/A history 2026-01-12 12:36:29 +02:00
Yiorgis Gozadinos
5a1cb253e5
Message history 2026-01-12 12:36:29 +02:00
Yiorgis Gozadinos
9f1d82daa8
Citations 2026-01-12 12:36:29 +02:00
Yiorgis Gozadinos
f8964f6efc
Basic frontend 2026-01-12 12:36:28 +02:00
Yiorgis Gozadinos
050ea8df70
Basic backend for app 2026-01-12 12:36:28 +02:00
Yiorgis Gozadinos
747acd6f6a
vb 2026-01-12 12:08:28 +02:00
Yiorgis Gozadinos
061095e855
Remove obsolete migrations 2026-01-12 12:07:49 +02:00
Yiorgis Gozadinos
119ca84da8
Merge pull request #227 from ggozad/feat/docling-doc-compression
Fix large storage overflow by changing docling_document_json (str) to docling_document (compressed bytes)
2026-01-12 11:57:31 +02:00
Yiorgis Gozadinos
29c682d8eb
Add migration 2026-01-11 11:40:57 +02:00
Yiorgis Gozadinos
9c4fd08c47
Update client serialization for compressed docling documents 2026-01-11 09:23:56 +02:00
Yiorgis Gozadinos
be715dd456
Revert docling core back to 1.8.0 for compatibility with docling-serve 2026-01-11 09:23:19 +02:00
Yiorgis Gozadinos
afc5b02285
Update schema to store docling_document as compressed bytes
- Change DocumentRecord.docling_document_json (str) to docling_document (bytes)
- Use large_binary Arrow type (64-bit offsets) to avoid 2GB column limit
- Decompress in Document.get_docling_document() using gzip
- Update DocumentRepository field mappings
2026-01-10 15:17:27 +02:00
Yiorgis Gozadinos
face6f95bc
compression utils 2026-01-10 15:11:58 +02:00
Yiorgis Gozadinos
09acc845c0
Update docling, docling-core, lancedb deps 2026-01-10 15:06:56 +02:00
Yiorgis Gozadinos
ff1f19bb08
vb 2026-01-08 18:50:39 +02:00
Yiorgis Gozadinos
bd1ea4e8a2
Merge pull request #222 from ggozad/fix/image-in-chunks
Fix base64 images leaking into expanded search results
2026-01-08 18:49:32 +02:00
Yiorgis Gozadinos
3603b00513
Fix base64 images leaking into expanded search results
PictureItem.export_to_markdown() defaults to ImageRefMode.EMBEDDED,
which embeds base64 image data. When _extract_item_text() called this
method during context expansion, base64 data would leak into results.

Now explicitly handles PictureItem with ImageRefMode.PLACEHOLDER to
prevent base64 while still including VLM descriptions and captions.
2026-01-08 18:42:58 +02:00
Yiorgis Gozadinos
4cae2cf26a
Clarify docs about using a VLM with docling-serve 2026-01-08 10:57:52 +02:00
Yiorgis Gozadinos
92f0fd956f
vb 2026-01-08 10:25:25 +02:00
Yiorgis Gozadinos
af380c8e52
Merge pull request #220 from ggozad/fix/openai-non-reasoning-model
Fix non-reasoning OpenAI models that do not support reasoning config
2026-01-08 10:18:50 +02:00
Yiorgis Gozadinos
1616891c40
Fix non-reasoning OpenAI models that do not support reasoning config 2026-01-07 18:47:31 +02:00
Yiorgis Gozadinos
ef187767ed
vb 2026-01-07 11:20:48 +02:00
Yiorgis Gozadinos
8939b8a313
Merge pull request #218 from ggozad/feat/vlm-image-handling
Support for VLM-based picture description for embedded images
2026-01-07 11:18:23 +02:00
Yiorgis Gozadinos
2f413e937b
Add VLM model to download_models() 2026-01-07 11:09:24 +02:00
Yiorgis Gozadinos
248eed8d7b
Update docs 2026-01-07 11:09:23 +02:00
Yiorgis Gozadinos
f861664656
Extract shared DoclingServeClient for async workflow & use it for VLM support, chunking, converting 2026-01-07 11:09:23 +02:00
Yiorgis Gozadinos
3b7fd440cc
End-to-end test for VLM annotation 2026-01-07 11:09:23 +02:00
Yiorgis Gozadinos
769e0d4746
Update picture description prompt, move its config under prompts 2026-01-07 11:09:23 +02:00
Yiorgis Gozadinos
3c4116397d
Add VLM picture description support for image handling 2026-01-07 11:09:22 +02:00