test_breaker_isolates_sources snapshotted job status after a fixed 0.2s
sleep and failed on a loaded runner while two good jobs were still in
flight. Poll for the drained jobs under a 5s deadline.
Assert the paused source's jobs are unattempted: the queued-uri assertion
alone passes with source isolation disabled, since a retried job returns
to QUEUED.
Make the MCP server read-only and multi-database; replace ask_question and analyze with execute_code; ship a Claude Code and Codex plugin with the haiku-rag skill
Every failure reaches the client as an MCP error carrying its message;
mask_error_details is set off explicitly, since FastMCP also reads it from
the environment. The masking goes, and with it the filter pre-check that
ran a count before every filtered call and the UnknownDatabaseError
translations that existed only to survive it. Explicit domain errors stay.
analysis.code_timeout was enforced only in _run_on_loop, the bridge for
database-bound reads; metadata.json, the cached JSONL files and in-code
search() and list_documents() never looked at the clock, and Monty's
watchdog counts compute only. Every host call now checks the deadline
before it starts.
Host errors keep their message for every caller: the masking added for the
MCP server goes, and the sandbox is one path for the capability and the
server alike.
The execute_code description gains the toc.json node shape, the patterns
the evals put into the analysis instructions (one list_documents call to
map a title to an id, files carry no source, toc before search for a known
document, doc_item_refs are items self_refs, chunk ids join files and are
not citations) and the sandbox's read-only, no-network, time-limit and
output facts, which the analysis instructions now state too. The skill
gains pictures as images, image search when offered, and citing chunk
metadata locators. docs/mcp.md lists the interpreter's limits under Code.
pydantic-monty>=0.0.23 brings collections, itertools, functools,
dataclasses, function decorators and str.format into the sandbox; every
layer names the same modules, and a test imports them. Monty now caps host
callbacks per checkout at 1000 by default; the sandbox raises it out of
reach, since the time budgets govern.
The Claude Code plugin moves to plugins/haiku-rag/ and gains a Codex
manifest at .codex-plugin/plugin.json; both share one .mcp.json and one
haiku-rag Agent Skill. .agents/plugins/marketplace.json serves Codex, the
Claude marketplace points at the new path, and scripts/bump_version.py
bumps both manifests. The skill declares its compatibility.
build_toc stored item positions, while the sandbox instructions describe
item_range as a slice into items.jsonl; a gap in positions pulled the next
heading into a section. Ranges are now indices into the position-ordered
items, and get_document_section slices by index too. docs/mcp.md names the
tools that take sources.
In Claude Code the client is the model, so the server no longer runs one.
execute_code runs a Python program per call in the analysis sandbox over
the selected documents and returns what it printed; the sandbox is created
and closed per call so Monty's cumulative budget and a frozen mount never
outlive a program. --no-agents goes with the two tools, and format_citations
in haiku.rag.utils goes with its only caller.
The sandbox exposes chunk metadata to code: chunk_meta on search results,
metadata on list_documents rows and in metadata.json, and chunks.jsonl per
document. A host-side failure inside a program, a document read or an
in-code search raising, reaches the program by exception type only and is
logged with its traceback. recovery_hint moves to haiku.rag.sandbox.
Closes#604.
Search results carry text and image blocks only. Claude Code and the
Agent SDK do not forward text blocks when structuredContent is present
and Desktop forwards both, so sending both either hid the rendering or
doubled it. The invalid-filter error keeps the engine's diagnosis and
lists our columns instead of lance's internals. format_citations no
longer repeats the URI of an untitled document.
Refs #599
Both search tools pass their results through HaikuRAG.expand_context, as
every other consumer of search results already did, so a client reads
the hit in its section rather than the chunk that matched. The rendering
gains an opt-in include_chunk_meta that shows the metadata stored with
the matched chunk beyond haiku.rag's structural keys, labelled as the
matched chunk's because an expanded passage spans several chunks and
only the anchor's metadata survives expansion. The capabilities' rendering
is unchanged.
Refs #599
fastmcp>=4.0.2,<5.0.0. Protocol types are snake_case (read_only_hint,
mime_type, input_schema, server_info); the camelCase names warn and go
in fastmcp 5, so the next major is an explicit upgrade. The client
defaults to the sessionless protocol, where initialize_result is None,
so the tests read client.instructions and client.server_info, which both
protocol modes populate. The server answers either mode.
claude-plugin/ holds the plugin manifest, the server configuration
(haiku-rag mcp --stdio, the configuration decides the database) and a
skill that says when to reach for the knowledge base and how to move
from a search result to a document, a section, an answer or a
computation. A repo-root marketplace manifest makes
`claude plugin marketplace add ggozad/haiku.rag` work. The skill
pre-approves every tool the server registers, and a test keeps the two
in step.
The manifest carries the package version, which bump_version.py now
rewrites: a versioned plugin updates only on a bump, so the installed
skill stays in step with the haiku-rag release the user has.
Refs #599
haiku-rag mcp --no-agents reaches create_mcp_server(agents=False):
ask_question and analyze are not registered and the instructions drop
the line describing them. qa.model always has a default, so a server
without a usable model cannot be detected from configuration; the flag
is how an operator says so.
Refs #599
search_documents and search_documents_by_image return a ToolResult: the
format_for_agent rendering with rank, Document ID and Collection so the
text alone drives the document tools; one ImageContent per distinct
picture, labelled with its result; and the SearchResult list without
image_data as structured content. format_for_agent gains an opt-in
include_document_id, so the capabilities' rendering is unchanged.
collect_pictures is the one place pictures are deduplicated and
validated for both wire formats.
Refs #599
build_toc moves from the sandbox into haiku.rag.context; the sandbox
keeps its toc.json unchanged. get_document_outline returns the heading
tree with page numbers and get_document_section one section's text,
subsections included, both resolved in the database holding the
document. Chunk ids never leave the server. ask_question drops `cite`
and always appends its citations.
Refs #599
FastMCP masks unexpected exceptions and logs the traceback server-side,
so paths and provider URLs never cross the transport. Expected failures
raise ToolError with a message: unknown document, unknown collection,
invalid filter, invalid base64, and agent failures naming only the
exception type. No tool returns an empty value or an error string on
failure any more.
A filter is validated on its own before the read that would use it, with
a filtered count on one of the selected databases: that is the query
engine rejecting the filter and nothing else, so its message (columns and
the statement) can be forwarded, while a ValueError raised later in the
read stays masked and no database outside the selection is opened.
Refs #599
FastMCP gets instructions, the haiku.rag-slim version and a lifespan.
Every tool carries read-only ToolAnnotations with a title, a description
that says when to use it, and a description on every parameter. The
filter description lists the document columns from DocumentMetaRecord
and how to match metadata; filter is accepted by both search tools.
One strict base64 decoder serves every image parameter.
DocumentInfo carries metadata.
Refs #599
`haiku-rag mcp` passes covers_set=True and _covering no longer refuses a
scope over several databases. search_documents, search_documents_by_image,
ask_question and analyze take `sources`; get_document takes `source`;
DocumentInfo carries `source`. format_citations gains include_source, which
ask_question sets from covers_multiple so citations name their database
only when the server covers several.
Refs #599
add_document_from_file, add_document_from_url, add_document_from_text
and delete_document are gone. create_mcp_server and _covering lose
read_only; the client always opens with read_only=True, so the
--read-only flag before `mcp` is redundant and the docs, Dockerfiles and
compose example drop it.
Ingestion is `haiku-rag add`/`add-src`/`delete` and haiku-ingester. No
known consumer used the write tools; one stdio server per client window
made multi-writer the accidental default, and streamable HTTP has no auth.
The test_app stub drops a comment and __init__ that described run_mcp
constructing the client positionally; every path goes through _covering.
Refs #599
telemetry.configure() passes send_to_logfire="if-token-present", so on a
machine with a logfire token the CLI tests configured a live exporter:
spans and the periodic metrics export went to the developer's project, and
where the export fired while a cassette was open, VCR rejected it with
CannotOverwriteExistingCassetteException in record mode none.
A token resolves from LOGFIRE_TOKEN or from logfire_credentials.json under
LOGFIRE_CREDENTIALS_DIR (default .logfire, relative to the working
directory), so an authenticated developer uploads with the variable unset.
conftest drops the variable and points credentials discovery at an empty
temporary directory, alongside the existing HAIKU_RAG_CONFIG_PATH setup and
before any haiku.rag import. The explicit send_to_logfire argument beats
LOGFIRE_SEND_TO_LOGFIRE, so denying the token is what disables the exporter.
test_split_and_merge_matches_single_pass resolved tests/data/doclaynet.pdf
relative to the working directory, so it failed with FileNotFoundError for
any invocation outside the repository root. conftest.py resolves the same
file as Path(__file__).parent / "data"; this was the only site that did not.
Replaces gpt-oss on ModelConfig, qa.model and processing.title_model, and
ministral-3 on the picture-description model. qa.model.vision follows the
model and is now true.
enable_thinking was gated on the gpt-oss name, so it did nothing for
qwen3.8. With title_model's max_tokens of 100 the reasoning consumed the
whole budget and title generation returned an empty string. The mapping
now applies to any ollama model via reasoning_effort(): false sends
"none", true sends "high". Measured on qwen3.8:27b-mlx, "low" does not
disable thinking and "none" does; gpt-oss is the inverse, its template
has no "none" level, so it keeps "low".
Picture description bypasses get_model -- docling posts the request
itself from a params dict -- so the flag was inert on that path too.
vlm_api_params() carries reasoning_effort into both converters' request
bodies. At max_tokens 200 the description survived either way, but the
switch cut completion tokens from 141 to 45.
test_search_tool_skips_binary_content_when_qa_model_is_text_only asserted
the vision default rather than setting it; it now configures vision=False
itself.
docs/benchmarks.md keeps ministral-3: those are recorded measurements.
A configured or default database that does not exist raises
SourceUnavailableError with the way to create it and without its location.
An unknown --db-name lists the databases there are, configured or default.
The CHANGELOG records Store.db_path, SingleDatabaseSession(ref, config),
the empty lancedb.uri migration and the error-type change.
An empty lancedb.uri, which init-config used to emit, is told to remove the
key; only a non-empty value is told the databases spelling. haiku-rag --db
and haiku-ingester --db turn a path with no stem into typer.BadParameter.
HAIKU_RAG_DB and DB_PATH are gone: a capability covers what the
configuration places or the db_path it is given, and the app backend and
the AG-UI example load their configuration as the CLI does. The compose
files point HAIKU_RAG_CONFIG_PATH at the mounted haiku.rag.yaml, which
places the database at /data where DB_VOLUME is mounted; the backend
refuses a configured set since it serves one database. The chat scopes a
selection by source only over a set and names databases on filter rows
only across several. Docstrings, docs and test fixtures stop describing an
unnamed database; every database a search, listing or citation reports
carries a name.
DatabaseRef is a name and a location. The configuration places databases
through lancedb.databases alone; with none configured the default is the
entry haiku.rag under storage.data_dir, selectable like any other.
lancedb.uri is removed, and a config carrying it fails to load with the
replacement spelled out. A path passed from Python is valid where the
configuration places nothing and raises AmbiguousDatabaseError beside
lancedb.databases; haiku-rag --db and haiku-ingester --db construct the
scope directly, so a human's override keeps working. Every database
answers to a name, and a database given as a path keeps its own errors.
Store, connect_lancedb, gather_database_info and run_doctor take a
location, a path or a URI, and classify it with ConnectionMode.of.
SingleDatabaseSession owns the resolved DatabaseRef and passes its
location down. This removes DatabaseRef.connection(), default_db_path,
the placeholder path for URI-backed databases and the per-database
config copies, so the configuration a client holds is the one the caller
gave it. The chat hands its capabilities the scope it opened along with
the client it lends, and the v0.58.0 migration no longer checks local
free disk for a database behind a URI.
A picture deduplicated within a burst must still reach the model once in
its own question and, when cited, again via the capsule re-fetch after
compaction; uncited it is dropped. Two wire tests pin both paths. The
harness builds the rag capability with defer_loading=False: a deferred
evidence capability the model has not loaded presents an empty record,
so compaction computes boundary 0 and rewrites nothing.
qa.md describes max_searches in units.
Sibling searches emitted in one response overlap heavily (40.6% of
returned chunk slots on Glimmer ORB fan-out cases). A result whose
rendered evidence a sibling already showed keeps its rank slot but
collapses to a reference line, and a picture attaches once per response
keyed on (source, document_id, self_ref). Equivalence is the
format_for_agent rendering at neutral rank/total plus picture keys,
bucketed under the qualified chunk id, so another database's copy or a
different expansion of the same anchor formats in full.
Search state now commits only after formatting and image construction
succeed: a raising image build no longer leaves results citable that
the model never saw, notes evidence for them, or suppresses a later
sibling.
Searches a model emits in one response share a budget unit, up to
FREE_SIBLINGS_PER_ROUND (3) per unit; sequential searches pay one unit
each, as before. Grouping keys on RunContext.run_step, which pydantic-ai
increments once per model request. A budget-rejected round fails all its
remaining siblings, and tracking resets per run.
Glimmer opens most questions with a burst of ~3 rephrasings in a single
response (95.8% of its three-search ORB cases are one-response bursts),
spending 3 of 5 searches before reading anything. Pass rate at 3 calls
equals 1 call, so a burst is priced as one probe.
The IVF probe count was the one vector search parameter with no
setting, and it is what bounds recall on a large indexed corpus:
partitions grow with the corpus, so a fixed probe count covers less of
it, and vector_refine_factor can only re-score what the probes
returned. Defaults to 20, matching lance, so search is unchanged.
Also corrects the create-index docs, which predate the measurement:
optimize() covers new chunks, so a rebuild is about retraining
centroids rather than reaching unindexed rows.
Retrieval scores are each database's own rank arithmetic; the databases
in a selection share an embedder, so similarity in that one space is the
signal comparable across databases by construction. Measured product to
product against score ordering: +8.3 to +16.6pp recall@5 across five
cells on two corpora, flat in collection count and corpus shape where
score ordering dips with both, closing roughly 60% of the gap to a
reranker; order-sensitivity residual 0.00pp in every cell. Exact ties
collapse from 51-81% of candidates to under 1%. Full-text-only searches
keep retrieval-score order, having no query vector. The vector column
already travels with every search result, so the similarity costs no
additional transfer; per-chunk embeddings are materialized only for the
federated path that reads them.
Rank interleaving guarantees every database slots regardless of content;
on domain-split collections it allocates no better than chance and costs
4.7pp recall@5 at four collections against score ordering (7.1pp at
eight). Hybrid scores are each database's own vector/FTS rank agreement,
which carries across databases; equal scores resolve by within-database
rank, and only a tie on both falls to configured order, leaving
permutation sensitivity at 0.02-0.26pp. Fused results carry the
candidate's own retrieval score, so the context-expansion re-sort
preserves fused order.
Disjoint corpora give every database's rank-r candidate the same RRF
score, and the stable sort resolved those ties to lancedb.databases
declaration order, discarding the retrieval scores entirely. Ties now
break on the raw retrieval score, which is uncalibrated across indexes
but only ever orders candidates within one rank tier: the databases in
a fusion share an embedder and ran the same search type, and it can
never lift a candidate above another rank. Hybrid per-database scores
are themselves rank-derived, so exact agreement still ties and keeps
configured order, deterministically. The n > limit depth quota is
unchanged, pending the retrieval eval.
Benchmarks state that every published number is exact brute-force kNN;
no benchmark database carries a vector index. The Vector Indexing
section gains the measured with/without IVF_PQ comparison (hotpotqa,
orb_multimodal_nemotron, frames: free to ~121k chunks, 0.0044 MAP at
426k, ~30 s / 4 GB build) and corrects the re-indexing story: optimize()
folds new chunks into the index as delta parts via auto_vacuum, so a
rebuild is about retraining centroids, not covering new rows.
