haiku.rag.config exported two configuration instances: the lazy _config
behind get_config/set_config, and Config, loaded at import time. Nothing
linked them, and eleven signatures captured Config as a default argument,
so set_config could not reach the factories, the client, the store or the
MCP server. reranking/base.py went further and snapshotted the configured
reranker name into a class attribute at import.
Config is removed. Internal defaults are config: AppConfig | None = None,
resolved through get_config() per call. RerankerBase._model is None and
CohereReranker takes its model name as an argument, like every other
reranker.
The suite patched attributes on Config while production read the instance
get_config() returns, a different object, so those patches were no-ops
waiting to happen. They now go through get_config().
Four call sites repeated lock, snapshot, try/except, restore. Each used
restore_table_versions, which restores in _tables() order — documents
first, contradicting RESTORE_TABLE_ORDER — and each caught Exception, so
a cancellation mid-write skipped rollback and left the earlier table
writes committed.
Store.write_transaction() holds the lock, snapshots under it, and rolls
back through _rollback_to_snapshot: RESTORE_TABLE_ORDER, shielded from
cancellation, absorbed cancellation re-delivered, rollback failure raised
with the original as cause. The two single-table update_meta sites keep
the bare lock.
The batch documents write moves inside the guarded body; it was outside
the try, so a failure there was never rolled back. Auto-vacuum is
scheduled after the transaction rather than inside it.
restore_table_versions is removed; those four sites were its only callers.
The setting was documented but did not exist on DoclingServeConfig, so it
was silently dropped, and DoclingServeClient.from_config never forwarded
the timeout parameter it already accepted. The per-request timeout was
therefore pinned at the constructor default of 300s with no way to change
it. Add the field, forward it, and reject a non-positive value.
Also parametrize over the checked-in *.yaml.example files and validate
each through AppConfig, so an example that no longer loads fails a test
rather than a user's first run.
create_document_from_source resolves a fetcher per call and never closed
it, so every one-shot URL or WebDAV ingest leaked an httpx connection
pool. Close it, but only when we built it: resolve_adhoc_fetcher returns
a caller-supplied source when one matches the URI, and the ingester keeps
those open across jobs.
Directory ingestion also yielded symlinked files resolving outside the
directory it was given. rglob does not recurse into symlinked
directories, so a symlinked file was the only way out of the tree; skip
those, as FSSource.discover already does.
The chunk repository docstring named client._ensure_chunks_embedded,
which does not exist.
FULL rebuild deleted a document before re-ingesting it from its URI, and
the handler around that logged and continued. A 404, a timeout or any
conversion error therefore removed the document permanently.
Deleting after a successful create is not an alternative:
create_document_from_source resolves the same URI to the existing
document and updates it in place, so a trailing delete would remove the
freshly rebuilt row.
Refresh in place instead. create_document_from_source takes an internal
force flag that skips the revision and MD5 short-circuits, so an
unchanged source is still re-converted, re-chunked and re-embedded into
the existing document, and the document id survives a rebuild.
A failed refresh now falls through to the stored-content path rather
than skipping the document: FULL recreates the chunks table before the
loop, so skipping left the document present but unsearchable until the
next rebuild.
The pending-batch flush moves out of the try. A failed flush is a lost
write and should abort the rebuild, not be logged and skipped.
Collapsing the caption text into `get_pictures_grouped` served the enrichment
path, which uses it, but the multimodal reranker discards the second return value
while still paying to read the column. That is the widest fan-out in the codebase,
`limit * 10` candidates, and it previously projected self_ref and picture_data
alone.
`with_text` is opt-in and off by default, so the cheap projection is what a caller
gets unless it asks for more. The reranker test asserts the projection as well as
the query count, since a count alone would not notice the column coming back.
Splitting the assembly into a passthrough pass and an expandable pass reordered
equal-scored results: the score sort that follows is stable, so the order results
arrive in is the tiebreak. Results are assembled in document_groups order again,
after the batched fetch rather than around it.
Also ports the caption negative cases the removed single-document test carried: a
table's caption and an ordinary text reference map to no picture.
`resolve_refs`, `get_items_in_range`, `get_caption_picture_refs` and
`get_all_items_grouped` have no callers left: the grouped equivalents serve every
path that used them. `get_all_items_grouped` had none even before this branch.
Tests whose subject was a removed method go with it. Tests that only used one to
fetch a fixture now use the grouped call, so what they assert is unchanged.
`_attach_picture_data` fetched picture bytes once per document, over the
`limit * 10` candidates reranking asks for, so it was the per-document fetch with
the most candidates behind it. It now issues one query however many documents the
candidates span: one for ten documents, as for one.
Removes `get_text_for_refs`, whose only caller now gets the text back with the
bytes from `get_pictures_grouped`.
`test_client_search_include_images_false_skips_lookup` returned no search
results, so asserting the picture accessor went uncalled held whatever the code
did. It now returns a picture-carrying result, making "did not fetch" the
assertion rather than "had nothing to fetch".
`expand_with_items` fetched its own inputs per document: one query to resolve
refs to positions, one for the window of items around them. A result set spanning
N documents cost 2N queries, which was 10 of the 18 measured for a limit=5 search
on a remote object-store corpus.
`expand_context` now does both fetches once for every document it is expanding,
and `expand_with_items` takes the positions and items it needs. Two queries for
one document, and two for five.
Each document keeps its own inclusive window in `get_items_in_ranges`. Positions
repeat across documents, so a shared range would splice one document's items into
another's context.
`_populate_image_data` ran its stages once per result document, so a result set
spanning N documents cost 4N `document_items` queries. Measured on a remote
object-store corpus, a limit=5 search with expansion was 18 queries, 16 of them
against `document_items`.
The stages now run once each across every document, and flat in document count:
two queries for the dependent caption-to-picture mapping when results ranked on a
caption, one for the picture bytes. Two queries for a picture-ref result set,
three at most.
Picture text comes back with the bytes rather than from a second query, since it
is on the same rows.
Predicates are per document, `(document_id = 'a' AND self_ref IN (…)) OR (…)`,
rather than `self_ref IN (union)`. self_ref and position values repeat across
documents, so a union predicate would return other documents' rows: for
picture_data that fetches blobs nobody asked for, and it can hand one document
another document's picture.
`client.ask` and `client.analyze` built their capability from a db_path, so the
capability opened a second connection to the database the client already had
open, once per call.
Ownership is now explicit rather than inferred. `rag` stays the connection the
capability opened and must close; `borrowed_rag` is a caller's, which
`_ensure_rag` prefers and `_close` never touches. Two fields rather than a flag,
so closing a borrowed connection is not expressible.
`for_run` still clears `rag` per run, since a run owns what it opens. It leaves
`borrowed_rag` alone: that connection belongs to the caller and outlives the run.
All ten tool bodies opened their own `HaikuRAG`, so every tool call paid a
connection open and, on object storage, refetched the index the previous call had
just cached. The client is now opened once, lazily so that calling a tool
function directly still works, and eagerly from the lifespan so an unopenable
database fails startup instead of every call. Teardown clears the cached client
in a finally, since `_lifespan_manager` can be re-entered and would otherwise
hand out a closed connection, including when the close itself fails.
`delete_document` no longer opens its own connection with `skip_validation=True`.
Keeping it separate broke consistency once connections became long-lived: the
delete committed on one connection while reads served from another, which with a
30s consistency interval showed the deleted document as still present. A
connection always sees its own writes, so sharing one is what makes delete
visible to the next read.
So the server no longer opts out of embedding-config validation. Drift that
validation rejects now fails MCP startup, where before the server started and
only `delete_document` worked while every read returned empty. Same-dimension
identity drift still starts a read-only server, matching every other read verb.
Delete under drift is now a CLI operation; CLAUDE.md and the CHANGELOG record it.
Every `Store` built its own connection with its own caches and discarded them on
close, so the index a vector query loads was refetched by the next connection.
On object storage that first fetch dominates: measured on a ~500k-chunk 2560-dim
corpus over a ~200ms link, the first query cost ~41s and the second ~3s, and a
new connection reusing the session cost ~7s instead of ~47s.
`connect_lancedb` now passes a process-wide session, keyed on the configured
cache sizes so a caller asking for different sizes gets its own.
Also sets `read_consistency_interval`, defaulting to 30s. It was None, meaning a
connection never re-checked for other processes' writes. Per-call connections hid
that; a shared session makes connections long-lived enough for a reader to go
stale against the ingester.
All three settings reject negatives at the config boundary. A negative cache size
raises OverflowError and a negative interval panics inside Lance, so neither is
catchable further in. Zero stays valid for both: no cache, and check on every
read.
The routing tests now assert the kwargs they care about rather than the full call
signature, since every connection carries the two new kwargs.
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.
`_init_tables` left `chunks.id`, `chunks.document_id` and `documents.id`
unindexed, so those lookups scanned the column. On object storage that is
network I/O per query, on paths that run per document: citation lookup,
delete-by-document, the re-ingest merge, and every dedup probe.
Declare the index set per table in `index_specs()` and apply it through
`ensure_indexes()`, which skips a column only when it is already indexed with
the declared type. Both halves of that are load-bearing. Skipping is required
because `create_index(replace=True)` rebuilds an identical index, writing a new
index and a new table version and orphaning the old files until the next vacuum.
Comparing the type is required because column coverage alone would let a
wrong-typed index stand, and a BTree on `label` silently loses the
low-cardinality equality lookup the Bitmap is there for.
Columns not declared for a table are left alone, so an externally created index
such as a vector index on `chunks` survives.
`_init_tables`, `recreate_embeddings_table`, `ChunkRepository.delete_all` and
`DocumentRepository.delete_all` now all route through it instead of repeating
their own subsets.
Also recreate `document_items` from `get_document_items_arrow_schema()` in
`DocumentRepository.delete_all`, which was using the LanceModel and so returned
`picture_data` as 32-bit `binary`.
Existing databases are unchanged; the migration follows separately.
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.
_store_documents_with_chunks embedded each document's chunks in its own
embed_chunks call; chunks missing embeddings are now flattened across the
whole batch, embedded in one pass honoring embeddings.batch_size, and
assigned back positionally.
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.
A cite call naming one good id and one mangled one registered the good one and
then asked the model to cite again. A model that mangles ids obeys, mangles
again, and the run dies on output retries with the answer lost: observed at 22
consecutive cite calls against gemma4-26b, with the citation policy registered
to press for a declaration. The branch is only reachable once something has
registered, so the answer already has grounding and there is nothing to ask for.
Resumption was inferred from the transcript, and no shape says it. A missing
prompt is how UI adapters ask their first question as much as how pydantic-ai
resumes one, so every AG-UI host failed on its first message. Reading the tail
instead moved the error rather than fixing it: a settled structured answer ends
with an output tool's return, indistinguishable from results delivered to a
question still in progress, so following questions inherited the first one's
identity.
`CapabilityEvidenceRecord.in_progress` is now the authority. `begin_question`
sets it, `after_run` clears it unless the run is only pausing for deferred work,
and a run that raised never reaches `after_run`, which is what leaves an
interrupted question claimable. The history is consulted only to catch a host
that dropped the state of a question the model is unmistakably still owed.
The documentation claimed a conversation that has never cited anything is not enforced,
which reads as though a first question could answer from fresh evidence without
declaring it. Enforcement needs neither condition to hold: no evidence outcome in this
question, and nothing cited earlier.
The structured-output test accepted a redirect or a recorded violation, so a break in
ending detection would still have passed on the backstop alone. The cite tool is
available in that scenario, so it asserts the redirect.
An output tool call is a `ToolCallPart` like any other, and treating every tool call as
intermediate meant a model could search, skip citing, emit its structured answer and
finish with neither a redirect nor a record. A response ends the question when it
carries no tool calls, or when one of its calls names an output tool.
Some endings are not visible from a single response — a host running
`end_strategy="early"` can finish on text beside a function call — so `after_run` is
the backstop: it cannot ask the model for anything by then, but it records a question
that reached the end of its run undeclared. That also covers a question that was asked
once, ignored, and finished anyway, which previously returned early on the redirect
marker and went unrecorded.
The capability documentation and the changelog said a question that gathered no
evidence is left alone. That describes neither the code nor the intent: enforcement
applies wherever there is something to declare, which includes a follow-up that reuses
evidence cited earlier without searching again.
`CitationPolicyCapability` makes the single enforcement decision, whatever mix of
evidence capabilities is registered: two of them must not each demand a citation for
one answer. It decides in `after_model_request`, when a response carries no tool calls
and the question can still be redirected. An explicitly ungrounded answer is a
declaration and is left alone; a question that gathered no evidence is left alone too,
read from the ledger rather than from a searches dict that a new question clears. When
the cite tool is already withdrawn the question is recorded in
`CitationPolicyState.violations` instead of pointing the model at a tool that is gone.
Registering it is the only switch. `DiscoveredEvidence` and discovery move to
`capabilities.evidence` so both optional capabilities share them, and `cite_available`
joins `evidence_tool_names` as public for the same reason.
Measured on Qwen3.6-35B over two arms of 37 questions, 29 of them unanswerable from
the corpus: explicit ungrounded declarations rose from 23 to 26, grounded answers to
unanswerable questions fell from 4 to 2, answerable questions stayed at 8 of 8, and
one redirect fired in the whole arm. Reading every answer found no invented grounding.
`_compact_old_tool_returns`, `PRIOR_TURN_NOTICE` and `turn_start` leave
`RAGCapabilityBase`, along with its `wrap_model_request` hook. The evidence
capabilities now retrieve and validate, and nothing else. Registering the compaction
capability is what rewrites a request; leaving it out sends the transcript untouched,
which was never a choice a host could make before.
The boundary is the recorded question identity rather than message shape, so a
resumption compacts what lies below the question in progress instead of switching
compaction off for the whole run. The newest earlier evidence return carries the
capsule and every other becomes a receipt, so one capsule exists by construction and
every return stays paired with its call.
Pictures of cited evidence are fetched through the capability that retrieved them and
re-attached beside the capsule with fresh labels. Ownership of a picture on the wire
requires the machine tag we write and an image directly after it, since neither
position nor prose is proof: several tools' results can arrive in one request, and a
user quoting our wording above their own picture had it removed. A picture that cannot
be fetched or decoded is emitted with neither its image nor its label.
The chat TUI and the example backend register the compactor, being multi-turn.
`client.ask`, `client.analyze` and the MCP tools do not: a single-shot question has
nothing earlier to compact.
A question identity is unset until a run establishes it. Testing whether the
record exists cannot stand in for that: a host with no state to send seeds a
default record, and a default record is truthy, so a resumption missing its real
state would have proceeded as question zero.
Every way of recording a message count now refuses one behind what is already
stored, through one shared check: a new identity, an evidence outcome and a
declaration, each against the newest identity, evidence epoch and declaration
epoch. Identities and epochs are only comparable while the conversation grows,
and `before_model_request` results are assigned back onto history, so that is a
constraint on the host rather than a guarantee of the framework. Left unchecked,
an evidence outcome moving backwards freezes every later declaration as stale,
and a declaration moving backwards replaces a newer one with an older one and
revives the answer it grounded.
The searches, citations and executions of a question in progress survive a
resumption. Clearing them cost the results the model was still answering from: a
citation afterwards recorded no provenance and could not resolve against the
expanded result it had seen, falling through to a database lookup.
A code execution counts as evidence when it succeeded or printed something. A
raised error with an empty stdout grounds nothing.
`CapabilityEvidenceRecord` holds the relationships a transcript cannot express:
which chunks a capability retrieved, which it cited, in which questions, and at
which point in the conversation. RAG and analysis each own one in their own state
namespace. Nothing is co-written: the host's state is JSON storage, so a shared
record would be overwritten by whichever capability synced last, and merging
happens in transient per-request views instead.
Both clocks are derived from the conversation rather than counted locally, so
every participant computes the same values without sharing a counter. Question
identity is the message count when the question arrived; epoch is the message
count at an outcome. Epochs are therefore globally comparable, which is what
lets `citation_status` require a declaration to follow the newest evidence of
every capability, and what makes equal epochs mean one request.
A declaration is written only after `resolve_citations` succeeds, so a call
naming only unresolvable ids is not a citation. Status is derived, never stored,
so refs and status cannot contradict.
Resuming a question requires the host to carry the capability state from the run
being resumed. Without it the identity of the question in progress is unknowable,
and adopting the current message count would relabel that question as a new one
and judge every declaration in it against the wrong identity.
Nothing reads the records yet and no wire behaviour changes.
A run carrying no prompt is continuing a question rather than asking one:
pydantic-ai resumes that way for deferred tool results, interruptions and
suspended responses. len(ctx.messages) then counts the live question's own
messages, so its search result was replaced by the earlier-question notice and
the model was asked to answer with the evidence removed. Reproduced: resuming
with an in-flight history left a notice where the only evidence was.
Switch compaction off for the whole run when ctx.prompt is None. The absence of
a prompt is the signal rather than the message layout — the resume shapes differ
from each other, and deriving the boundary from layout is what broke this to
begin with.
ToolReturn.content reaches the model as a user-role message and the pictures
arrive bare, so nothing connects a figure to the chunk it came from:
BinaryContent.identifier does not survive serialization to the vision API, and
the captions in the result text correlate only by position.
Precede each picture with its position, source chunk id and self_ref.
build_binary_parts_from_results becomes build_image_content_from_results and
returns the labels interleaved with the pictures, so both attachment sites emit
them the same way.
This does not stop a model narrating retrieved pictures as user-supplied.
Measured on gemma4-26b with a single note ahead of the batch, and again with
per-image labels: it quotes the label and still says the user provided them.
The message role wins over its text.
_compact_old_tool_returns ran in before_model_request, whose result core
assigns back onto ctx.state.message_history, so the trim reached
all_messages() and every host that persists a thread. It now runs in
wrap_model_request, which operates on a detached list: the model sees the
trimmed history, the host keeps what it retrieved.
Content attached to a ToolReturn arrives as its own UserPromptPart in the
same ModelRequest as the ToolReturnPart, so the turn-boundary scan read an
image-bearing search result as a new user turn and discarded evidence
retrieved earlier in the same turn. _is_user_turn() now requires a request
with no tool returns.
Page images on a replaced return stay. Dropping them with their text bounds
context growth, but a follow-up about a figure already shown ("what colour
is that box?") carries no terms that could retrieve it again: measured on
gemma4-26b against two ORB figures, removing the image turned both answers
into "I cannot find enough information", and keeping it answers correctly.
Bounding that growth needs to preserve cited figures, which is a separate
change.
The replacement notice no longer claims citations remain in state;
_clear_invocation_state has cleared them by then.
mxbai-rerank-base-v2 ships a Sigmoid activation and evaluates it in bf16, so
every strongly-relevant candidate rounds to exactly 1.0. Ties then leave the
order to the stable sort, which preserves the incoming hybrid ranking: on 100
t2_finqa retrieval cases the reranker scored MAP 0.661 against 0.659 with no
reranker at all, and 0.742 once the scores separate.
Ask the model for logits and apply the sigmoid here, where it runs in float64.
Scores stay 0-1, matching the cohere, vllm and zeroentropy rerankers.
Also drop the remaining pyright references; the project type-checks with ty.
