Yiorgis Gozadinos
27ed0b3bb3
Send analysis to the sandbox when only its search budget is spent
...
The notice told the model to answer from what it had the moment
qa.max_searches ran out, while up to 15 code executions remained and
in-code search() does not count against that budget. It now names the
spent tool and points at whichever evidence tool still has budget,
falling back to answer-and-cite only when none do.
Also count RetryPromptPart in n_failed_tools: _cite rejects with
ModelRetry, so a run whose every cite attempt was refused reported zero
failures. And note that n_requests is the run's request count, which
tracks a capability's own budget only while it stays loaded.
2026-07-30 18:17:26 +03:00
Yiorgis Gozadinos
b62d6e920b
Drop budget_spent from eval attributes
...
It was n_rejected_searches > 0 recorded next to the integer it derived
from, and the name overclaimed: analysis_execute_code also raises
ToolFailed when execute_count exceeds max_executions, which the flag
never saw. Callers can compare the counters directly.
2026-07-30 15:57:48 +03:00
Yiorgis Gozadinos
721acbcf38
Address review on PR #524
...
- _budget_notice no longer names the cite tool after prepare_tools has
withdrawn it; the post-grace state gets the plain no-tools text back.
- Split search-budget rejections from any failed tool call: the code tool
raises ToolFailed for every error in model-written Python, so
budget_spent was true for a ZeroDivisionError.
- docs/capabilities/rag.md described the old single-turn removal.
- Drop the rationale clause from the CHANGELOG entry.
2026-07-30 15:50:06 +03:00
Yiorgis Gozadinos
a528ab912f
Record per-case retrieval diagnostics in evaluations
...
Count tool traffic from the message history, where refused and repeated
calls stay visible, unlike state.searches which is keyed by query.
2026-07-30 14:56:17 +03:00
Yiorgis Gozadinos
ebd1fc7f3e
vb
2026-07-29 09:06:55 +03:00
Yiorgis Gozadinos
044da7ae99
Open the eval database read-only outside population
...
Retrieval and QA only read from the database, but the benchmark opened it
writable, where an embedder identity differing from the stored one aborts
instead of warning. Running a pre-built database against a different
serving stack then needed a `rebuild --set-embedder` first.
Correct the debug-evals skill alongside it: the pydantic-ai span names are
`execute_tool {tool_name}` and `invoke_agent agent`, targets are
`{rag,analysis}-capability`, and no `skill_model` metadata key exists.
2026-07-28 11:36:27 +03:00
Yiorgis Gozadinos
94e03d777a
vb
2026-07-25 15:11:54 +03:00
Yiorgis Gozadinos
acbd66afbd
vb
2026-07-24 17:32:20 +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
9deb1f2bd4
replace haiku.skills with native Pydantic AI capabilities
2026-07-24 15:26:17 +03:00
Yiorgis Gozadinos
ec78641cc4
vb
2026-07-24 15:15:20 +03:00
Yiorgis Gozadinos
9f49bdde6e
vb
2026-07-23 14:28:23 +03:00
Yiorgis Gozadinos
294ced57cd
vb
2026-07-23 10:45:48 +03:00
Yiorgis Gozadinos
4f0fb7322e
vb
2026-07-22 13:27:35 +03:00
Yiorgis Gozadinos
11303aa714
Document hotpotqa benchmark results and finalize the reference config
2026-07-17 16:28:16 +03:00
Yiorgis Gozadinos
0e8fa551f3
Restore hotpotqa evaluation dataset
2026-07-17 16:25:27 +03:00
Yiorgis Gozadinos
b5c0ea630c
vb
2026-07-16 13:50:40 +03:00
Yiorgis Gozadinos
5dd310a9b7
vb
2026-07-14 16:06:50 +03:00
Yiorgis Gozadinos
144900d385
Remove the mxbai reranking provider
2026-07-14 11:09:55 +03:00
Yiorgis Gozadinos
5c41f8a0ed
vb
2026-07-10 13:38:24 +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
da0748916b
vb
2026-07-09 15:58:58 +03:00
Yiorgis Gozadinos
a1cf405cae
Rename document_meta identity column document_id to id
2026-07-08 16:32:55 +03:00
Yiorgis Gozadinos
9a9bea86d3
vb
2026-07-03 13:35:10 +03:00
Yiorgis Gozadinos
9a85632cf3
vb
2026-06-29 12:36:36 +03:00
Yiorgis Gozadinos
325f4517ba
Update t2_finqa config
2026-06-29 12:22:58 +03:00
Yiorgis Gozadinos
3cb229d2e0
Add t2_finqa pre-built evaluation database reference config and docs
2026-06-29 10:52:23 +03:00
Yiorgis Gozadinos
4d03b1e669
Add nemotron-vl multimodal eval database and reference configs
2026-06-29 10:14:53 +03:00
Yiorgis Gozadinos
b2abdfaffb
vb
2026-06-28 10:57:26 +03:00
Yiorgis Gozadinos
c7c3f75508
vb
2026-06-27 09:59:21 +03:00
Yiorgis Gozadinos
cd1a73b389
vb
2026-06-26 17:13:03 +03:00
Yiorgis Gozadinos
197d58a602
vb
2026-06-24 18:41:05 +03:00
Yiorgis Gozadinos
ab1775dbc6
vb
2026-06-24 16:45:19 +03:00
Yiorgis Gozadinos
65f1339aa0
vb
2026-06-23 16:59:00 +03:00
Yiorgis Gozadinos
0fcf91fcf7
vb
2026-06-22 15:18:35 +03:00
Yiorgis Gozadinos
5c4b37df89
vb
2026-06-19 11:12:45 +03:00
Yiorgis Gozadinos
99011755fb
vb
2026-06-16 16:48:36 +03:00
Yiorgis Gozadinos
df8af54298
rebase from main
2026-06-12 10:17:18 +03:00
Yiorgis Gozadinos
88696cc6f1
vb
2026-06-11 09:50:33 +03:00
Yiorgis Gozadinos
bdad908a60
vb
2026-06-09 15:39:14 +03:00
Yiorgis Gozadinos
46b8fa3fef
vb
2026-06-08 12:24:04 +03:00
Yiorgis Gozadinos
295570797c
Disable Logfire scrubbing in evaluations
2026-06-08 11:51:13 +03:00
Yiorgis Gozadinos
e9bd56467c
Add T²-RAGBench leaderboard submission exporter
2026-06-08 11:51:02 +03:00
Yiorgis Gozadinos
a3a73f1331
Add --filter-ids to run QA on a case-id subset
2026-06-08 09:36:01 +03:00
Yiorgis Gozadinos
1d7d540270
Match the numeric scale convention in Number-Match
2026-06-06 14:52:05 +03:00
Yiorgis Gozadinos
d5b0aafa2b
Add T²-RAGBench TAT-DQA subset; generalize subset layout
2026-06-06 14:52:05 +03:00
Yiorgis Gozadinos
de9731d5f3
Score Number-Match on the declared ANSWER line, magnitude and ×100 scale
2026-06-06 14:52:05 +03:00
Yiorgis Gozadinos
390deb4203
Normalize unicode signs
2026-06-06 14:52:04 +03:00
Yiorgis Gozadinos
80594cd38c
Handle percent/decimal convention in Number-Match
2026-06-06 14:52:04 +03:00
Yiorgis Gozadinos
c2a48b55c0
Add deterministic Number-Match QA scoring for T²-RAGBench
2026-06-06 14:52:04 +03:00
Yiorgis Gozadinos
9f1bd9940a
Add T²-RAGBench FinQA evaluation dataset
2026-06-06 14:52:04 +03:00
Yiorgis Gozadinos
916b84bff4
vb
2026-06-05 12:09:57 +03:00
Yiorgis Gozadinos
e70f271ac7
vb
2026-06-04 14:13:55 +03:00
Yiorgis Gozadinos
4eae86225e
vb
2026-06-03 14:26:58 +03:00
Yiorgis Gozadinos
eb74525084
Lift analysis-skill citation rate via SKILL.md tightening
2026-06-01 18:58:52 +03:00
Yiorgis Gozadinos
8dbcd4bd7a
vb, changelog
2026-06-01 18:54:22 +03:00
Yiorgis Gozadinos
e653c49fce
Remove unecessary Mean Reciprocal Rank metric
2026-06-01 10:40:52 +03:00
Yiorgis Gozadinos
00f2a40a60
Refresh benchmarks doc and remove unused eval datasets
2026-06-01 10:40:51 +03:00
Yiorgis Gozadinos
62be23d130
vb
2026-05-29 11:47:12 +03:00
Yiorgis Gozadinos
3828a91c37
vb
2026-05-27 16:23:43 +03:00
Yiorgis Gozadinos
14d9951016
vb
2026-05-21 14:32:13 +03:00
Yiorgis Gozadinos
392c74039b
bump pydantic-ai-slim to 1.100 and haiku.skills to 0.17
...
Migrate off two APIs slated for removal in pydantic-ai 2.0:
- Agent(tool_retries=, output_retries=) -> Agent(retries={"tools": ...,
"output": ...}) in the LLM-as-judge evaluator.
- Evaluator.evaluation_name class attribute -> overriding
get_default_evaluation_name() on the citation MRR / MAP evaluators.
haiku.skills 0.17.0 already migrated its internal AGUIAdapter,
MCPToolset and ProcessEventStream usage; no further changes needed on
our side beyond the pin bumps.
2026-05-21 13:20:54 +03:00
Yiorgis Gozadinos
98b0c068e8
vb
2026-05-20 14:34:33 +03:00
Yiorgis Gozadinos
f9392cee46
evaluations: judge model moves to config.evaluations.judge
2026-05-20 12:56:22 +03:00
Yiorgis Gozadinos
8c57a3ca99
Drop the multi-agent research workflow
2026-05-20 12:46:48 +03:00
Yiorgis Gozadinos
ceb645564c
Drop GEPA prompt optimization and --target qa from evaluations
2026-05-19 11:01:07 +03:00
Yiorgis Gozadinos
74dea65f21
analysis.model inherits qa.model when unset; per-skill vision gate
2026-05-18 16:49:43 +03:00
Yiorgis Gozadinos
1df6530d03
Bump pydantic-ai, prepare for 2.*
2026-05-18 15:14:29 +03:00
Yiorgis Gozadinos
97a7eb246c
vb
2026-05-14 16:03:33 +03:00
Yiorgis Gozadinos
862a209137
vb
2026-05-13 16:53:35 +03:00
Yiorgis Gozadinos
cc59dc7203
Use upload_large_folder for hf upload
2026-05-06 13:12:08 +03:00
Yiorgis Gozadinos
ce7201271f
split open_rag_bench dataset into orb_text and orb_multimodal variants
2026-05-06 12:49:03 +03:00
Yiorgis Gozadinos
a45d82f206
Update benchmark
2026-05-05 13:58:48 +03:00
Yiorgis Gozadinos
5c3a864a0b
rename open_rag_bench eval DBs to text/multimodal variants
2026-05-05 12:55:06 +03:00
Yiorgis Gozadinos
75c8e83515
Allow uri override on create_document_from_source
2026-04-30 16:36:35 +03:00
Yiorgis Gozadinos
6a77ce92a9
extract picture bytes to document_items.picture_data at ingest, strip them from the docling_document blob, and add 0.45.0 migration to backfill existing
...
databases
2026-04-30 11:26:37 +03:00
Yiorgis Gozadinos
29bc42514e
vb
2026-04-29 13:45:36 +03:00
Yiorgis Gozadinos
8962987ff9
pin judge model to ollama:qwen3.6
2026-04-29 12:24:06 +03:00
Yiorgis Gozadinos
5e31c15907
docs
2026-04-28 14:44:27 +03:00
Yiorgis Gozadinos
7bfb818d77
remove dataset-specific system prompts
2026-04-28 14:33:25 +03:00
Yiorgis Gozadinos
7d288c525e
citation retrieval scoring
2026-04-28 12:46:16 +03:00
Yiorgis Gozadinos
d00befd0c4
benchmark RAG and analysis skills via --target
2026-04-28 12:09:19 +03:00
Yiorgis Gozadinos
6807f082ec
vb
2026-04-25 09:35:47 +03:00
Yiorgis Gozadinos
81ba248483
vb
2026-04-24 17:47:57 +03:00
Yiorgis Gozadinos
4f058bf484
vb
2026-04-22 17:00:03 +03:00
Yiorgis Gozadinos
bb9e7b28ca
vb
2026-04-22 16:18:44 +03:00
Yiorgis Gozadinos
8a23108fbd
vb
2026-04-20 16:36:32 +03:00
Yiorgis Gozadinos
ff8ad0879c
fix context expansion: respect section boundaries, remove max_context_items
2026-04-20 15:18:16 +03:00
Yiorgis Gozadinos
1049586469
vb
2026-04-17 16:24:25 +03:00
Yiorgis Gozadinos
4f16714430
prevent stale files from accumulating on HF during uploads.
2026-04-16 13:45:24 +03:00
Yiorgis Gozadinos
e92f4774fa
Improve judge prompt
2026-04-16 12:11:54 +03:00
Yiorgis Gozadinos
b6113bf8ab
replace fixed-radius expansion with section-bounded algorithm
...
Context expansion is now automatic and structure-aware. For structured
documents, expands within the section containing the match. For sections
that exceed the budget or are too small, expands item-by-item outward
skipping noise labels. Unstructured documents use budget-based outward
expansion. Results sorted by relevance score.
2026-04-16 12:11:53 +03:00
Yiorgis Gozadinos
0d6df472af
bump ag-ui protocol deps, fix duplicate React key in chat
2026-04-14 12:50:25 +03:00
Yiorgis Gozadinos
dcde86a5f1
vb
2026-04-09 15:33:19 +03:00
Yiorgis Gozadinos
62a436dd9d
vb
2026-04-08 14:04:32 +03:00
Yiorgis Gozadinos
3d30840e9d
vb
2026-04-07 16:36:27 +03:00
Yiorgis Gozadinos
e4dc60c4f1
Update core dependencies
2026-04-03 10:14:27 +03:00
Yiorgis Gozadinos
5c4164d224
vb
2026-04-01 13:32:56 +03:00
