build(deps-dev): bump the python-dependencies group across 1 directory with 10 updates

[dependabot]

Updates the requirements on mcp, mypy, fonttools, protobuf, ruff, transformers, haystack-ai, pyspark, pymdown-extensions and yasbd-lib to permit the latest version.
Updates mcp to 2.3.0

Release notes

Sourced from mcp's releases.

v2.3.0

pip install -U mcp. Docs: https://py.sdk.modelcontextprotocol.io/

Mostly fixes, plus three new options. A few things behave differently, so skim these first:

Behaviour changes

httpx2>=2.10.0 is now required (#3600)

• It was >=2.5.0. The new max_sse_event_size option needs it.
• Nothing to do unless you pin httpx2 below 2.10.

A tool with an invalid x-mcp-header annotation fails at registration (#3620)

• @mcp.tool(), add_tool and Tool.from_function raise InvalidSignature, naming the tool and the problem.
• Until now the server started, and 2026-07-28 clients silently dropped the tool from their listing.
• Refused: anything other than a plain str, int or bool parameter (so also str | None, float, lists and enums), a header name that isn't a valid token, and two names that differ only by case.
• For an optional header parameter, give the schema directly: Annotated[str | None, WithJsonSchema({"type": "string", "x-mcp-header": "Region"})] = None.

Empty _meta and params are no longer sent (#3628)

• On 2025-11-25 and earlier connections, 2.x sent "_meta": {} on every request. Some servers reject that. It is now left out, as in v1.
• ping and list requests without a cursor go out with no params member.
• On the receiving side ctx.meta is None rather than {}, and middleware sees ctx.params as None for a request without params.
• 2026-07-28 connections are unchanged.

initialize leaves out experimental when none is configured (#3614)

• It used to send "experimental": {}. server/discover already left it out.
• Client code reading capabilities.experimental on a legacy connection should handle None.

Mcp-Param-* validation looks the tool up by name (#3630)

• MCPServer no longer runs tools/list for every tools/call, so middleware no longer sees that extra request.
• The registered schema is what gets checked. Middleware that filters or rewrites tools/list no longer affects it.

An interactive OAuth login no longer counts against request timeouts (#3635)

• The timeout pauses while OAuthClientProvider waits on redirect_handler and callback_handler.
• This fixes Client(mode="auto") settling on 2025-11-25 when the login took longer than 10 seconds.
• A request timeout no longer ends a login nobody finishes. Put a limit inside callback_handler if you need one.

New

• max_sse_event_size= on streamable_http_client and StreamableHttpParameters. The default stays 1 MiB per SSE event; raise it, or pass None, for larger tool results (#3600).
• MCPServer(subscriptions=False) stops serving subscriptions/listen and advertises listChanged and subscribe as false (#3626).
• Server(get_tool_input_schema=...) lets a low-level server supply a tool's schema for header validation without running its tools/list handler (#3630).
• Client.call_tool re-lists the tools and retries once after a HeaderMismatch (-32020) rejection (#3627).

Fixes

• ctx: Context[AppState] works on prompts and resource templates, not only on tools (#3624).
• An explicit "structuredContent": null is checked against the output schema instead of being treated as missing (#3621).
• A progress_callback that raises no longer fails the call on an in-process Client(server) (#3623).
• Client OpenTelemetry spans record JSON-RPC error responses. With mode="auto", connecting to a server without server/discover now shows one ERROR span for the probe (#3629).
• stdio_client resolves the executable off the event loop on Windows (#3510).

... (truncated)

Commits

• 2118f14 docs: refresh translations for recent English changes (#3636)
• ed9b2d6 Stop counting an interactive OAuth login against request timeouts (#3635)
• d5cebd1 Keep inline-snapshot disabled when pytest runs in a terminal (#3634)
• c15566c Link What's new to the Header parameters page (#3632)
• 4d29994 Let a newer Deploy Docs run cancel the one in progress (#3633)
• 0acea60 Bump urllib3 from 2.7.0 to 2.8.0 (#3607)
• c54075c Look the tool schema up by name for Mcp-Param-* validation instead of running...
• 9afccae Retry a tool call once after a HeaderMismatch rejection (#3627)
• cafa33b Record JSON-RPC error responses on the client OpenTelemetry span (#3629)
• 0b2fd3e Omit an empty _meta and empty params from outbound requests (#3628)
• Additional commits viewable in compare view

Updates mypy from 2.3.0 to 2.4.0

Changelog

Sourced from mypy's changelog.

Mypy Release Notes

Next Release

Mypy 2.4

We've just uploaded mypy 2.4.0 to the Python Package Index (PyPI). Mypy is a static type checker for Python. This release includes new features, performance improvements and bug fixes. You can install it as follows:

python3 -m pip install -U mypy

You can read the full documentation for this release on Read the Docs.

Python 3.15 Support

Mypy 2.4 supports running on Python 3.15 and type checking most Python 3.15 features. This includes the new builtin sentinel type for sentinel values (PEP 661), which mypy now supports (see below for details). Lazy imports (PEP 810) and unpacking in comprehensions (PEP 798) are also supported. Closed TypedDicts (PEP 728) have been supported since mypy 2.2, but the extra_items TypedDict argument, also introduced in PEP 728, is still unsupported. Support for extra_items will be added in a future mypy release.

Native Parser Enabled by Default

Mypy now uses the new native parser by default. It's based on the Ruff parser, and it's significantly faster than the legacy parser, which uses the stdlib ast module. The native parser also has other benefits:

• You can target newer Python versions and use recent Python syntax even when running mypy on an older Python version. For example, you can use --python-version 3.15 when running mypy on Python 3.14.
• Stub files can use syntax that is newer than the target Python version. For example, stubs can use the PEP 695 generic class syntax (class Box[T]: ...) when running on or targeting Python 3.10.
• Parallel type checking requires the native parser.

The legacy parser is still available through --no-native-parser, or native_parser = False in the config file (native_parser = false under [tool.mypy] in pyproject.toml). We are planning to remove the legacy parser in early 2027. If you run into a problem with the native parser, please report it on the issue tracker.

Unlike the legacy parser, the native parser doesn't support type comments for variables defined by for and with statements. These type comments are silently ignored, and the types of the variables are

... (truncated)

Commits

• 935ba8a Bump version to 2.4.0
• 334a678 Add link for PR 22066 in mypy 2.4 changelog (#22082)
• fe69660 Finalize mypy 2.4 changelog (#22080)
• 541fd87 Various mypy 2.4 changelog updates (#22077)
• b231182 Update native parser docs (#22078)
• 04fe42c Initial Changelog for Mypy 2.4 (#22071)
• 0617494 Quick fix for non UTF-8 encodings with new parser (#22066)
• b2f9b60 Bump librt to 0.16.0 (#22067)
• 9d798dc Sync typeshed (#22054)
• 0861bb6 Reduce max protocol stack size to 20 (#22053)
• Additional commits viewable in compare view

Updates fonttools from 4.61.1 to 4.65.0

Release notes

Sourced from fonttools's releases.

4.65.0

• [glyf] Add __iter__, items and values methods to the glyf table to make it more dict-like (#4156).
• [feaLib] Escape the anonymous block tag when scanning for its terminator, so tags containing regex metacharacters are matched literally (#4167).
• [varLib] Strip directory components from <variable-font name="..."/> when deriving the output filename in the varLib command line, so a designspace cannot write outside the output directory (#4168).
• [feaLib] Fix tracking of the current script and language across redundant script statements. Rules following a script statement that names the first declared language system no longer end up under the DFLT script, and a script statement naming the already-current script still narrows the language systems and terminates the current lookup while leaving the lookupflag alone, matching makeotf (#1824, #2522, #4169).
• [varLib.interpolatable] Escape glyph names in the HTML report (#4172).
• [otlLib] Fix overflow handling when building contextual lookups: offset overflows now raise OTLOffsetOverflowError instead of AttributeError so another contextual format can be tried (regression from #3439). When all formats overflow, split the ruleset in halves until it fits (#4171).

4.64.0

• [feaLib] Fix name-table parsing for multibyte Mac encodings (#1196, #4092).
• [ttProgram] Also indent TrueType assembly following IDEF[ ], like function definitions (#4093).
• [subset] Keep East Asian spacing palt by default (#4094).
• [subset] Bug fix for MATH table in which constructions for glyphs that are only added during MATH closure were kept (#4096).
• [ufoLib] Make glyph-to-group construction accessible outside of lookup function (#4102).
• [glyf] Use reverse glyph map for O(1) __setitem__ membership (#4103).
• [ttLib] Fix fixLookupOverFlows() reporting success when it had not promoted any lookup to Extension, masking unresolvable overflows.
• [ttLib] Add support for TrueType Collection version 2 (#4100).
• [ttLib] Pin a single head.modified timestamp across TTCollection.save (#4111).
• [ttLib] Give an actionable error when LookupList overflow is unrecoverable (#4109).
• [ttLib] Add support for the AAT bitmap tables bhed, bdat, bloc, variants of head, EBDT, EBLC used in legacy Apple bitmap-only fonts (#4115).
• [ttLib] Check OS/2 fsSelection/macStyle consistency against bhed as well as head (#4118, #4119).
• [misc.roundTools] Add types and documentation (#4123).
• [varLib.instancer] Instance the BASE table (#4137).
• [varLib.instancer] Fix Private-dict vsindex handling in instantiateCFF2 (#4129, #4132).
• [varLib.instancer] Fix crash instancing CFF2 fonts without a VariationStore (#4130, #4131).
• [sfnt] Raise TTLibError instead of AssertionError or struct.error when reading a font truncated within the table directory or a table entry (#4147, #4149).
• [misc.xmlWriter] Escape the ]]> terminator inside CDATA sections, so an SVG document containing it can no longer smuggle markup past a TTX round trip (#4139).
• [varLib.instancer] Implement avar2 partial-instancing: the avar version 2 ItemVariationStore is adjusted so that remaining axes behave the same after limiting the designspace (#4045).
• [feaLib] Add shorthand for the value at the default location in a variable scalar: (100 wght=900:120) means (wght=400:100 wght=900:120) when the wght default is 400 (#4024).
• [cmap] Raise TTLibError for a truncated or out-of-bounds cmap subtable header (#4151).
• [designspaceLib] Reject conflicting duplicate inputs in axis maps instead of silently keeping the last one (#4153).
• [designspaceLib] Read an empty <lib> element as an empty lib instead of raising IndexError (#4142, #4144).
• [colorLib] Raise a legible error when a COLRv0 layer, or a COLRv1 PaintGlyph or PaintColrGlyph, references a glyph missing from the glyphMap, instead of failing obscurely later (#2629, #4141).
• [cmap] Don't drop subtables in unsupported formats when compiling or dumping a font read from binary (#4136).
• [ttLib] Implement splitSinglePos so GPOS lookup type 1 offset overflows can be recovered by splitting the subtable (#4091, #4108).
• [cmap] Round-trip empty Macintosh format 2 subtables (#3663, #4117).
• [glyf] Raise TTLibError instead of RecursionError when recalcBounds() hits a composite-component reference cycle (#3899, #4116).
• [svgLib] Fix crash parsing an SVG path with consecutive closepath commands (Z Z) (#4122).
• [ttLib] Fix DefaultTable type annotations (#4126).
• [ttLib] Add support for the EBSC (Embedded Bitmap Scaling) table (#4113).
• [svgLib] Suppress spurious close segments caused by floating-point drift in relative path commands (#3860, #4127).
• [qu2cu] Fix TypeError in the Cython-compiled build when Qu2CuPen passes tuple splines (#4160).
• [mort] Add semantic decompilation, TTX, and compilation support for rearrangement, contextual-substitution, ligature, and insertion subtables (#4158, #4159, #4161).
• [svgLib] Start a new subpath at the just-closed subpath's initial point when a drawto command follows a closepath, per SVG spec (#4154, #4155).
• [misc.filesystem] SECURITY Reject paths that resolve outside the filesystem root: a malicious UFO could read arbitrary files via .. components in contents.plist, and a crafted .ufoz could create files outside its temporary mirror (#4124).
• [ttLib] SECURITY Sanitise glyph names used as filenames in EBDT/CBDT ttx -z extfile export, preventing arbitrary file writes from untrusted fonts (#4128).
• [misc.etree] SECURITY Don't resolve external XML entities in XMLParser when lxml is used, preventing XXE file disclosure on lxml < 5.0 (#4145).
• [subset] Fully prune VARC auxiliary data: collect and remap variation indices referenced by condition tables when subsetting the MultiVarStore, and drop the AxisIndicesList, ConditionList, and MultiVarStore when they end up empty (#4162).

4.63.0

... (truncated)

Changelog

Sourced from fonttools's changelog.

4.65.0 (released 2026-09-10)

• [glyf] Add __iter__, items and values methods to the glyf table to make it more dict-like (#4156).
• [feaLib] Escape the anonymous block tag when scanning for its terminator, so tags containing regex metacharacters are matched literally (#4167).
• [varLib] Strip directory components from <variable-font name="..."/> when deriving the output filename in the varLib command line, so a designspace cannot write outside the output directory (#4168).
• [feaLib] Fix tracking of the current script and language across redundant script statements. Rules following a script statement that names the first declared language system no longer end up under the DFLT script, and a script statement naming the already-current script still narrows the language systems and terminates the current lookup while leaving the lookupflag alone, matching makeotf (#1824, #2522, #4169).
• [varLib.interpolatable] Escape glyph names in the HTML report (#4172).
• [otlLib] Fix overflow handling when building contextual lookups: offset overflows now raise OTLOffsetOverflowError instead of AttributeError so another contextual format can be tried (regression from #3439). When all formats overflow, split the ruleset in halves until it fits (#4171).

4.64.0 (released 2026-08-31)

• [feaLib] Fix name-table parsing for multibyte Mac encodings (#1196, #4092).
• [ttProgram] Also indent TrueType assembly following IDEF[ ], like function definitions (#4093).
• [subset] Keep East Asian spacing palt by default (#4094).
• [subset] Bug fix for MATH table in which constructions for glyphs that are only added during MATH closure were kept (#4096).
• [ufoLib] Make glyph-to-group construction accessible outside of lookup function (#4102).
• [glyf] Use reverse glyph map for O(1) __setitem__ membership (#4103).
• [ttLib] Fix fixLookupOverFlows() reporting success when it had not promoted any lookup to Extension, masking unresolvable overflows.
• [ttLib] Add support for TrueType Collection version 2 (#4100).
• [ttLib] Pin a single head.modified timestamp across TTCollection.save (#4111).
• [ttLib] Give an actionable error when LookupList overflow is unrecoverable (#4109).
• [ttLib] Add support for the AAT bitmap tables bhed, bdat, bloc, variants of head, EBDT, EBLC used in legacy Apple bitmap-only fonts (#4115).
• [ttLib] Check OS/2 fsSelection/macStyle consistency against bhed as well as head (#4118, #4119).
• [misc.roundTools] Add types and documentation (#4123).
• [varLib.instancer] Instance the BASE table (#4137).
• [varLib.instancer] Fix Private-dict vsindex handling in instantiateCFF2 (#4129, #4132).
• [varLib.instancer] Fix crash instancing CFF2 fonts without a VariationStore (#4130, #4131).

... (truncated)

Commits

• 656f89d Release 4.65.0
• 1643950 Update NEWS.rst [skip ci]
• fd0f083 Merge pull request #4177 from fonttools/sbix-to-otsvg-snippet
• 9c625d4 Merge pull request #4171 from fonttools/fix-contextual-size-overflow
• 5f04c38 Add sbix to OT-SVG conversion snippet
• 48a0cf9 Merge pull request #4172 from insaf021/interpolatable-html-escape
• d30db24 Merge pull request #4169 from cmyr/fealib-script-tracking-afdko
• ac342d7 [varLib.interpolatable] escape glyph name in HTML report
• 2a93de2 Split contextual rulesets when all formats overflow
• e6bb5fe Fix overflow reporting when sizing contextual lookups
• Additional commits viewable in compare view

Updates protobuf to 7.36.2

Updates ruff from 0.15.22 to 0.17.0

Release notes

Sourced from ruff's releases.

0.17.0

Release Notes

Released on 2026-10-09.

The executables in our macOS and Windows release archives and ruff wheels are now code-signed. macOS executables are signed with an Apple Developer ID certificate and notarized by Apple. Windows executables have timestamped Authenticode signatures from Azure Artifact Signing. This enables verification of the release publisher and binary integrity, supports publisher-based allowlisting, and should reduce security warnings and antivirus false positives.

Breaking changes

• Update the default and latest Python versions for 3.15 (#28792)

  Ruff now defaults to Python 3.11 instead of 3.10 when no Python version is configured through target-version or requires-python. When checking for syntax errors without a configured Python version, Ruff now defaults to Python 3.15 instead of 3.14.

• Update the default rule set (#28786)

  Several of the flake8-datetimez rules (DTZ001, DTZ005, DTZ006, DTZ007, DTZ011, DTZ012, and DTZ901) are no longer enabled by default, while undefined-local-with-nested-import-star-usage (F406), which corresponds to a syntax error, is now enabled by default.

• Update Rust crate quick-junit to 0.8.0 (#27295)

  JUnit output now uses a skipped attribute instead of disabled on <testsuite> elements and includes a skipped attribute on the root <testsuites> element.

• [flake8-import-conventions] Add datetime as dt as a conventional alias (ICN001) (#28790)

• Support Unicode dummy variable names (#28722)

  The default lint.dummy-variable-rgx now recognizes underscore-prefixed Unicode names, such as _次, as dummy variables.

• Update to Unicode 17 (#21229, #28784)

  Ruff now uses Unicode 17 data for identifier normalization and named character escapes ("\N{...}").

• Always show unsafe and display-only fixes in the CLI (#27810)

  The default full output format now shows unsafe fixes and suggestions requiring manual review, regardless of the unsafe-fixes setting. Actually applying unsafe fixes still requires explicit opt-in.

• Remove the Python dependency from conda-forge builds (conda-forge/ruff-feedstock#361)

  The conda-forge build no longer depends on Python, now supports linux-riscv64, win-arm64, and linux-ppc64le platforms, and now includes shell completions. However, no longer depending on Python means that python -m ruff and import ruff will no longer work. Use ruff directly from PATH instead. PyPI installations and those from the standalone installer are unaffected.

• Remove support for ruff-lsp (#28750)

  Support for ruff-lsp, the legacy Python language server deprecated in Ruff v0.9.5, has been removed. The Ruff VS Code extension now always uses the native language server; ruff.nativeServer is deprecated and ignored. See the migration guide.

Stabilization

The following rules have been stabilized and are no longer in preview:

• lazy-import-mismatch (TID254)
• lazy-import-immediately-resolved (TID255)
• os-path-commonprefix (RUF071)

... (truncated)

Changelog

Sourced from ruff's changelog.

0.17.0

Released on 2026-10-09.

The executables in our macOS and Windows release archives and ruff wheels are now code-signed. macOS executables are signed with an Apple Developer ID certificate and notarized by Apple. Windows executables have timestamped Authenticode signatures from Azure Artifact Signing. This enables verification of the release publisher and binary integrity, supports publisher-based allowlisting, and should reduce security warnings and antivirus false positives.

Breaking changes

• Update the default and latest Python versions for 3.15 (#28792)

  Ruff now defaults to Python 3.11 instead of 3.10 when no Python version is configured through target-version or requires-python. When checking for syntax errors without a configured Python version, Ruff now defaults to Python 3.15 instead of 3.14.

• Update the default rule set (#28786)

  Several of the flake8-datetimez rules (DTZ001, DTZ005, DTZ006, DTZ007, DTZ011, DTZ012, and DTZ901) are no longer enabled by default, while undefined-local-with-nested-import-star-usage (F406), which corresponds to a syntax error, is now enabled by default.

• Update Rust crate quick-junit to 0.8.0 (#27295)

  JUnit output now uses a skipped attribute instead of disabled on <testsuite> elements and includes a skipped attribute on the root <testsuites> element.

• [flake8-import-conventions] Add datetime as dt as a conventional alias (ICN001) (#28790)

• Support Unicode dummy variable names (#28722)

  The default lint.dummy-variable-rgx now recognizes underscore-prefixed Unicode names, such as _次, as dummy variables.

• Update to Unicode 17 (#21229, #28784)

  Ruff now uses Unicode 17 data for identifier normalization and named character escapes ("\N{...}").

• Always show unsafe and display-only fixes in the CLI (#27810)

... (truncated)

Commits

• 5b27c1c Bump version to 0.17.0 (#29259)
• 8f362bf Integrate vendored Unicode names into Ruff
• 88a50d6 Revert upstream Unicode loose matching
• 276832a Revert upstream Unicode alias normalization
• 26911ca Vendor unicode_names2 from upstream
• 27e3797 [ruff] Stabilize os-path-commonprefix (RUF071) (#28764)
• bbb52c1 [flake8-tidy-imports] Stabilize lazy-import-immediately-resolved (`TID255...
• 5aa9cbd [flake8-tidy-imports] Stabilize lazy-import-mismatch (TID254) (#28760)
• dba1bc8 [flake8-import-conventions] Add datetime as dt as a conventional alias (`...
• ef23c9c Update default and latest Python versions for 3.15 (#28792)
• Additional commits viewable in compare view

Updates transformers from 5.12.1 to 5.19.0

Release notes

Sourced from transformers's releases.

Release v5.19.0

New Model additions

EmbeddingGemma2

EmbeddingGemma 2 is a multimodal embedding model from Google built on the Gemma 4 architecture. It encodes text, images, audio, and video, individually or combined in one input, into a shared 768-dimensional vector space for cross-modal retrieval, semantic similarity, clustering, and classification. It uses Matryoshka Representation Learning, so embeddings can be truncated to 512, 256, or 128 dimensions. It also offers configurable visual and video token budgets, and unused vision or audio towers can be disabled at load time to save memory.

Links: Documentation

• Smthn smthn (#49364) by @​vasqu in #49364

Breaking changes

All MoE models whose routers compute logits now return them when output_router_logits=True, following the Qwen3-MoE pattern (a router_logits recorder on the base model, MoeModelOutputWithPast from the backbone, and a MoE causal LM output from the head), so code that relied on the previous outputs or their absence should read the router logits from these output classes.

• 🚨 Return router logits from every MoE model that computes them (#48920) by @​qgallouedec

Owlv2ForObjectDetection.embed_image_query now selects the query box with the highest objectness score, as in the original OWLv2 notebook, instead of the OWL-ViT heuristic, so image-guided query embeddings and detections may differ from earlier releases.

• 🚨 Select the OWLv2 image query by objectness (#49200) by @​qgallouedec

The "paged|" prefix for SDPA and flash attention implementations is deprecated, so users should set the regular attention implementation (e.g. sdpa or flash_attention_2) for continuous batching instead of paged|sdpa or paged|flash_attention_2.

• 🚨 Attention 🚨 Deprecate "paged|" prefix for SDPA and flash (#49112) by @​remi-or

The regular flash and SDPA attention functions (flash_attention.py, sdpa_attention.py) now support continuous batching directly, and "paged|..." implementations for these are redirected to them, while eager still requires the "paged|eager" prefix.

• 🚨 Attention 🚨 Make regular attention support CB (#49101) by @​remi-or

In continuous batching, the cache update for the index-based and block-table paths is now fused into a single call, which slightly changes the cache update function's behavior and affects any custom code that calls the separate update paths.

• 🚨 [CB] 🚨 Fuse update for index and block table path (#49088) by @​remi-or

Continuous batching internals changed in preparation for removing "paged": `max

• 🚨 [CB] 🚨 Little fixes before removing "paged" (#49069) by @​remi-or

Parallelization

Expert parallelism gains a token-dispatch implementation, selected via the new ep_dispatch_experts plan rule and now the default for Qwen3 MoE and Mellum, which removes the requirement that EP size equal TP size. The Trainer was also adapted to work with expert parallelism, and the docs now note that PEFT adapters support tensor parallelism. A CI-related fix for pipeline-parallel chart2table inference was also included.

• [distributed]: adapt Trainer to work Expert Parallel (#48873) by @​3outeille in #48873
• [distributed] Add expert-parallel token dispatch, default for Qwen3 MoE (#48865) by @​3outeille in #48865
• Fix pp chart2table inference (#49225) by @​zucchini-nlp in #49225
• [docs] TP for PEFT adapters (#49060) by @​stevhliu in #49060

Cache

... (truncated)

Commits

• c4c4605 v5.19.0
• 92cd495 Smthn smthn (#49364)
• a7e8ff2 Revert "[CI] Temporarily disable AMD scheduled CI trigger" (#49271) (#49362)
• 571c5cc ALM testx fIxing... (#49355)
• b4526a7 Fix DataCollatorForLanguageModeling ignoring seed=0 (#49349)
• 5a27d2d Fix explicit kernelization modes for evaluation models (#49339)
• 23c21ac fix(cohere_compass): add video to input_modalities (#49320) (#49323)
• f1d0d9f GPT-OSS: read the clamped SwiGLU's alpha and limit from the config (#49346)
• 9921402 [distributed] Default MoE ep_plans to token dispatch (#49160)
• 8625785 [distributed]: adapt Trainer to work Expert Parallel (#48873)
• Additional commits viewable in compare view

Updates haystack-ai to 3.3.0

Release notes

Sourced from haystack-ai's releases.

v3.3.0

⭐️ Highlight

⚡️ Faster SentenceWindowRetriever

SentenceWindowRetriever now queries the Document Store once per run or run_async call instead of once per retrieved document. Fewer calls mean lower latency and less load on the Document Store and outputs have not changed, so existing pipelines get faster without code changes.

⬆️ Upgrade Notes

• InMemoryDocumentStore.bm25_retrieval and InMemoryBM25Retriever with BM25L (the default) or BM25Plus now return only documents that contain at least one query term, with or without scale_score. Previously, documents without any query term were returned with a positive score and filled up top_k. As a result, retrieval can now return fewer than top_k documents, or none at all. The delta lower bound now applies only to query terms that occur in the document, as in the original BM25L/BM25+ definitions. So scores of matching documents are lower than before whenever the query has terms that a document does not contain. If you filter results with a fixed score threshold, re-check the threshold. BM25Okapi is not affected.

⚡️ Enhancement Notes

• TextCleaner.run now raises a clear TypeError when texts is not a list or any element is not a str, instead of failing later or producing unexpected results.

🔒 Security Notes

• Haystack now requires anyio>=4.14.2. anyio is installed transitively through httpx and openai; versions before 4.14.2 are affected by CVE-2026-63374 (GHSA-82r6-8w77-94w6).

🐛 Bug Fixes

• Fixed BM25 tokenization in InMemoryDocumentStore for Chinese, Japanese and Korean text. The default bm25_tokenization_regex now splits CJK characters into one token each, so bare-term queries can match words inside longer unspaced runs. Text is also NFC-normalized before tokenization, so composed and decomposed spellings produce the same tokens.

• Fix Pipeline.connect() raising TypeError: object of type 'ellipsis' has no len() when one of the sockets is annotated with Callable[..., T]. A Callable[..., T] now matches callables with any parameters, in both directions, as long as the return types are compatible.

• Fixed ChatPromptBuilder dropping content parts when the template is a list of ChatMessage objects. Only the first TextContent part was rendered, and every other part - additional texts, images, files -was removed from the rendered prompt without a warning. Template variables used in those dropped parts were not detected either, so they were not exposed as inputs. Now all TextContent parts are rendered and the remaining content parts are passed through unchanged.

• Fixed CompactionHook.close() and close_async() to release the token counter's resources as well as the compactor's. Previously, resources such as OpenAITokenCounter's HTTP client remained open after the hook or Agent was closed.

• ConfirmationHook no longer drops the messages that come after the last user or tool message when it rewrites the conversation, such as a system message added by another hook or an assistant answer followed by an on_exit reminder. Before, those messages were removed from the history the Agent sends to the model.

• DocumentToImageContent no longer raises a KeyError for the whole batch when a document points to a PDF page that cannot be converted, because the page is out of range or the PDF cannot be read. That document now gets None in image_contents, like other invalid documents, and the rest of the batch is still converted. The logged warnings now include the path of the PDF file.

• Fixed DOCXToDocument breaking a Markdown table when a cell contains a pipe or spans several paragraphs. A pipe was emitted as a column separator, and a cell's line break ended the row in the middle of it. Pipes are now escaped and line breaks inside a cell are collapsed to a space. The csv table format is unchanged: it already kept such cells intact by quoting them.

• Fixed DOCXToDocument dropping hyperlink addresses inside tables. With link_format set to markdown or plain, links in table cells were written as their display text only, while links in body paragraphs kept their address. Links in table cells are now formatted the same way, in both the markdown and csv table formats. The default link_format="none" output is unchanged.

• Fixed ConditionalRouter.from_dict mutating the caller's routes data in place: serialized output_type strings were deserialized directly inside the caller's dictionaries, so reusing the same serialized pipeline dict afterwards yielded already-deserialized type objects. from_dict now works on a copy and the caller's data is left untouched. BranchJoiner.from_dict received the same treatment.

• LinkContentFetcher no longer adds its timeout and follow_redirects defaults to the client_kwargs dictionary passed by the caller. The dictionary is now copied before the defaults are applied, so reusing one HTTP client configuration across components no longer leaks these defaults.

• Fix MetaFieldRanker to return no documents when missing_meta="drop" and all documents lack the ranking field or have a None value.

• MetaFieldRanker now treats a Document whose meta_field value is None the same as a Document that is missing the field, applying the missing_meta setting to it. Previously a single None value made sorting fail, so the ranker logged a warning and returned all Documents in their original order, ignoring missing_meta="drop" as well.

• MockTextEmbedder and MockDocumentEmbedder now accept non-positive dimension values when embedding or embedding_fn is provided, matching the documented behavior. The default deterministic embedding mode still requires a positive dimension.

• ChatMessage.from_openai_dict_format now accepts tool-call arguments that are already a dictionary instead of raising a TypeError. Some OpenAI-compatible servers send a parsed object rather than a JSON string.

• MetaFieldGroupingRanker now treats a group_by or subgroup_by value of None as missing, the same way it already treats sort_docs_by. Documents with None go to the end with the other ungrouped documents, instead of forming a group named "None" that also absorbed documents whose value was the string "None".

... (truncated)

Commits

• daa2d1f bump version to 3.3.0
• 05b616c bump version to 3.3.0-rc1
• f4e004c docs: sync Core Integrations API reference (monty) on Docusaurus (#13033)
• 2ae9acd fix: render every content part of ChatPro...

  Description has been truncated