benchmark.py was 1220 lines holding six unrelated jobs: populating a
database, running retrieval, running QA, resolving datasets, moving
databases to and from HuggingFace, and wiring the Typer CLI.
qa.py takes both QA runners with their live summary, refusal metrics and
target resolution. population.py takes populate_db and the batched ingest.
retrieval.py takes run_retrieval_benchmark. artifacts.py takes HF_REPO_ID and
the download/upload bodies. experiment.py takes DEFAULT_JUDGE_MODEL and
build_experiment_metadata, which retrieval and QA both record.
benchmark.py keeps the CLI at 258 lines: the Typer app, config and case-id
loading, dataset resolution, evaluate_dataset, and three commands whose
bodies are now a loop over specs. The module-level side effects stay with it
— load_dotenv before configure_telemetry, so credentials and LOGFIRE_TOKEN
are in the environment before telemetry and model setup read them — so no
importable module carries one.
Test patch targets follow the code. get_model, run_capability_question,
run_capability_conversation, set_eval_attribute and HaikuRAG are patched
inside moved code, so they move with it; run_qa_benchmark,
run_retrieval_benchmark and find_config_file stay patchable on
evaluations.benchmark because evaluate_dataset and _load_config still look
them up there.
One assertion got stronger: a QA test patched benchmark.HaikuRAG to prove the
QA path does not open its own client. qa.py has no HaikuRAG reference at all
now, so the test asserts that instead.
RAGState and AnalysisState each declared the same five fields, so the generic
base could not name them: StateT was bound to BaseModel, and every access
went through cast(Any, state), a getattr by string, or a loop clearing fields
by name so it could skip the one only AnalysisState has.
EvidenceState declares them once. RAGState adds nothing, AnalysisState adds
executions and overrides begin_invocation to clear them. StateT binds to
EvidenceState, which removes all ten casts and both state-shape getattrs; the
three getattr(ctx.deps, "state") probes stay, since those check a
host-supplied object rather than our own state.
discover_evidence reached into capability.state for two fields. It now asks
through evidence_record() and citation_index(), alongside the
evidence_tool_names() and cite_available accessors it already used. The eval
runner's _RagLikeState protocol and the chat app's getattr reads described
this shape from outside and are gone.
Compatibility is semantic JSON-object equivalence, not bytes: field names and
nesting are unchanged, so a dict stored by 0.75.0 loads and re-dumps equal,
but deriving from a shared base reorders AnalysisState's keys. Nothing
serializes, hashes or string-compares this state — every carry point
re-validates by key.
Adds the 0.75.0 upgrade, which brings a pre-existing database up to the index
set `_init_tables` now creates. It rewrites no table data, so unlike the earlier
data migrations its cost is the index builds alone, each of which reads the
column it indexes.
`ensure_indexes` ensures an index of the declared *type* covers each declared
column, rather than checking that the column is indexed at all. The distinction
is what makes it safe to run against a database of unknown provenance:
- A wrong-typed index no longer satisfies the check. A BTree on `label` covers
the column while losing the low-cardinality equality lookup the Bitmap is for.
- Nothing is dropped or converted away from. Two index types over one column can
be deliberate, serving different query shapes, so an index this function did
not declare survives even on a column it does. The one thing it overwrites is
an index at LanceDB's default name, `{column}_idx`, which is the name it
creates itself.
- A column already carrying the declared type is skipped, so a database with the
full set migrates instantly rather than re-sorting every indexed column.
- Undeclared columns are untouched, so a vector index on `chunks` survives.
It returns the columns it acted on, because a change is not always visible from
outside: adding a Bitmap beside an existing BTree leaves the column indexed
before and after.
The version bump to 0.75.0 is required, not incidental: `_set_initial_version`
stamps a new database with the installed package version, so a migration
numbered above it would be pending the moment the database was created.
`test_client_update_document_replaces_rows_with_bounded_versions` turns
auto_vacuum off. Indexing `documents` means a background vacuum now has an index
to maintain on that table, so `optimize()` writes a version where it previously
had nothing to do, and it landed inside the window the test measures. The
document update itself is still one version, so the bound stays exact.
IBM's MTRAG benchmark (ClapNQ domain, pinned repo SHA): retrieval with
Recall@k/nDCG@k against binary qrels, gold-prefix QA replaying reference
conversation prefixes as message history, and live-session replay
carrying the model's own answers and tool history across turns.
Corpus population gains a bounded, resumable batched ingest path.
ConversationInput case type with transcript rendering for the judge,
eligibility-aware citation scoring, refusal precision/recall via a
label-aware RefusalJudge, per-turn verdicts with judged-turn coverage,
and per-turn tool-traffic attributes counted from each turn's new
messages so the arrays survive prior-turn compaction.
No `DatasetSpec` declared `search_filter`, so `resolve_search_filter`
and the `--filter ""` clearing rule reconciled the flag against a
default that never existed. The flag alone covers the case. An empty
clause reaches `ChunkRepository.search`, which already treats it as
unfiltered.
Rename to `document_filter` throughout, matching
`run_capability_question`'s parameter and the metadata key that lands
in Logfire.
`_stub_spec` merges its overrides, so a test can override a loader
instead of rebuilding the whole spec.
The `--filter` examples used `uri LIKE '%arxiv%'`, which matches no
`orb_text` document: its URIs are bare arXiv ids such as `2407.01528v3`.
A clause that matches nothing scores MAP 0 rather than erroring, so the
example failed silently.
`await_args` is typed `_Call | None`, so subscripting it fails
`ty check`; `call_args` carries the same call for an AsyncMock.
The request-limit notice said only the cite tool remained available, but
chat registers rag and analysis in one agent, so exhausting analysis
claimed rag_search was gone too. Scoped to the capability's own tools.
The cite window was counted over every model request once loaded, so
turns spent on another capability expired it before the model was ever
placed where citing was the obvious move. Count only requests whose
preceding response called one of this capability's tools; engagement is
also the only thing that can loop, which is all the bound guards against.
Also: _count_tool_traffic returns a named tuple rather than four bare
ints, and counts failures only for this capability's tools, so host-tool
retries and output-validation retries no longer read as its failures.
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.
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.
- _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.
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.
