[dependabot]

Bumps the services-minor-patch group with 31 updates in the /services directory:

Package	From	To
duar-auth	1.5.1	1.5.2
confluent-kafka	2.13.0	2.16.0
dateparser	1.3.0	1.4.3
eventsourcing	9.5.2	9.5.6
gliner2	2.0.0	2.1.0
peft	0.20.0	0.21.2
fastapi	0.142.2	0.161.1
langchain	1.4.5	1.4.7
langchain-community	0.4.1	0.4.2
langchain-core	1.6.10	1.7.0
langchain-ollama	1.0.1	1.1.0
langfuse	3.14.5	3.15.0
pydantic	2.12.5	2.14.1
pydantic-settings	2.14.2	2.16.1
pymongo	4.18.3	4.19.0
pymupdf	1.27.1	1.28.2
python-dotenv	1.2.2	1.2.4
qdrant-client	1.16.2	1.20.0
rdkit	2025.9.5	2025.9.6
returns	0.26.0	0.29.0
transformers	5.20.0	5.20.2
uvicorn	0.40.0	0.54.0
langchain-openai	1.3.3	1.7.2
langchain-anthropic	1.4.7	1.7.6
langchain-google-genai	4.2.5	4.4.2
nvidia-cuda-nvcc-cu12	12.8.93	12.9.86
pytest	9.0.3	9.1.1
pytest-asyncio	1.3.0	1.4.0
pytest-cov	7.0.0	7.1.0
ruff	0.14.14	0.17.2
watchfiles	1.2.0	1.3.0

Updates duar-auth from 1.5.1 to 1.5.2

Updates confluent-kafka from 2.13.0 to 2.16.0

Release notes

Sourced from confluent-kafka's releases.

v2.16.0 is a feature release with the following features, fixes and enhancements:

Enhancements

• confluent_kafka now declares itself GIL-safe, enabling real multi-core parallelism on free-threaded CPython builds. See the Multithreading Guide for thread-safety details and free-threaded caveats. (#2347)
• Add Python 3.14t wheels (#2352)
• Add Linux s390x (IBM Z) wheels, including Python 3.14t (#2360)
• Producer close() now aborts any open transaction (#2347)
• Async IO Consumer's default worker pool size has been increased from 2 to 100 (#2347)
• SerializingProducer and DeserializingConsumer accept a serde builder when not passing a ready-made serde, through the new key.serializer.builder / value.serializer.builder and key.deserializer.builder / value.deserializer.builder configuration properties. Serdes built this way, and any Schema Registry client the builder created for them, are owned by the client and closed by its close(); ready-made serdes remain the application's. This API is experimental (#2364).
• Serdes that resolve subjects through the Schema Registry associated subject name strategy are now given the Kafka cluster id automatically. The id is resolved lazily, on the first subject lookup, so creating a client never waits on a broker, and concurrent lookups share a single cluster_id() call; until a broker has been reached the lookup raises a SerializationError naming subject.name.strategy.kafka.cluster.id, which can be set to supply the id explicitly. The new serde hooks (set_cluster_id_resolver(), close()) are experimental (#2364).
• New Producer.cluster_id(), Consumer.cluster_id() and AdminClient.cluster_id() (also on AIOProducer and AIOConsumer), returning the id of the cluster the client is connected to. This API is experimental (#2364).
• New asyncio clients AsyncSerializingProducer and AsyncDeserializingConsumer in confluent_kafka.aio, the counterparts of SerializingProducer and DeserializingConsumer built on AIOProducer / AIOConsumer, accepting the asyncio Schema Registry serdes and their builders. These classes are experimental (#2364).
• New Message.deserialized_key() and Message.deserialized_value(), which return the same objects as key() and value() but are typed with the deserialized types on a DeserializingConsumer. SerializingProducer and DeserializingConsumer are now generic in their key and value types. This API is experimental (#2364).
• Add support for saving Azure key version with DEK (#2306)
• Pass context when clients make KEK calls to DEK Registry (#2308)
• Schema Registry: add support for the DLQ (dead-letter-queue) rule action (DlqAction). When a rule fails, the record is teed to a configured DLQ topic and the original serialize/deserialize call still raises. With the default (global) RuleRegistry the DLQ is best-effort; set dlq.auto.flush=true or give the serde its own RuleRegistry (closable on shutdown) for durability.
• Add support for inline validation rules (#2326)
• Add Variant, Decimal, and Timestamp CEL functions (#2332)

Fixes

... (truncated)

Changelog

Sourced from confluent-kafka's changelog.

v2.16.0

v2.16.0 is a feature release with the following features, fixes and enhancements:

Enhancements

• confluent_kafka now declares itself GIL-safe, enabling real multi-core parallelism on free-threaded CPython builds. See the Multithreading Guide for thread-safety details and free-threaded caveats. (#2347)
• Add Python 3.14t wheels (#2352)
• Add Linux s390x (IBM Z) wheels, including Python 3.14t (#2360)
• Producer close() now aborts any open transaction (#2347)
• Async IO Consumer's default worker pool size has been increased from 2 to 100 (#2347)
• Add support for CPython free-threading (PEP 703) to enable true multi-core parallel execution in No-GIL environments (#2347)
• SerializingProducer and DeserializingConsumer accept a serde builder when not passing a ready-made serde, through the new key.serializer.builder / value.serializer.builder and key.deserializer.builder / value.deserializer.builder configuration properties. Serdes built this way, and any Schema Registry client the builder created for them, are owned by the client and closed by its close(); ready-made serdes remain the application's. This API is experimental (#2364).
• Serdes that resolve subjects through the Schema Registry associated subject name strategy are now given the Kafka cluster id automatically. The id is resolved lazily, on the first subject lookup, so creating a client never waits on a broker, and concurrent lookups share a single cluster_id() call; until a broker has been reached the lookup raises a SerializationError naming subject.name.strategy.kafka.cluster.id, which can be set to supply the id explicitly. The new serde hooks (set_cluster_id_resolver(), close()) are experimental (#2364).
• New Producer.cluster_id(), Consumer.cluster_id() and AdminClient.cluster_id() (also on AIOProducer and AIOConsumer), returning the id of the cluster the client is connected to. This API is experimental (#2364).
• New asyncio clients AsyncSerializingProducer and AsyncDeserializingConsumer in confluent_kafka.aio, the counterparts of SerializingProducer and DeserializingConsumer built on AIOProducer / AIOConsumer, accepting the asyncio Schema Registry serdes and their builders. These classes are experimental (#2364).
• New Message.deserialized_key() and Message.deserialized_value(), which return the same objects as key() and value() but are typed with the deserialized types on a DeserializingConsumer. SerializingProducer and DeserializingConsumer are now generic in their key and value types. This API is experimental (#2364).
• Add support for saving Azure key version with DEK (#2306)
• Pass context when clients make KEK calls to DEK Registry (#2308)
• Schema Registry: add support for the DLQ (dead-letter-queue) rule action (DlqAction). When a rule fails, the record is teed to a configured DLQ topic and the original serialize/deserialize call still raises. With the default (global) RuleRegistry the DLQ is best-effort; set dlq.auto.flush=true or give the serde its own RuleRegistry (closable on shutdown) for durability.
• Add support for inline validation rules (#2326)
• Add Variant, Decimal, and Timestamp CEL functions (#2332)

... (truncated)

Commits

• 4a1bcb0 Copy pyproject.toml into smoketest temp dir so asyncio_mode=auto applies (#2384)
• 16ba967 Fix Unicode errors in string serialization (#2379)
• 73a71f8 2.16.0 (#2380)
• 20614ae Integrate Schema Registry serdes into the Kafka clients: serde builders, clus...
• 7f2e72f Add s390x wheels to the build and verification matrix (#2360)
• 6a54a83 chore: normalize service.yml field names (#2193)
• c57c5b8 Fix Producer.purge() ignoring boolean flags on big-endian platforms (#2345)
• 89c99e0 On Windows, the tag isn't currently given when building librdkafka, that affe...
• cc1c970 v2.16.0rc2 (#2369)
• d245efd Fix crash on nullable array of $ref items in JSON Schema CSFLE (#2370)
• Additional commits viewable in compare view

Updates dateparser from 1.3.0 to 1.4.3

Release notes

Sourced from dateparser's releases.

1.4.3

Fixes:

• Make parsing thread-safe: parsing from several threads no longer raises an intermittent KeyError from the shared language caches, and a DATE_ORDER or RELATIVE_BASE value meant for one parse no longer leaks into the settings that other parses read, where it could make them return a wrong date (#1346)
• Do not share the language detector and the detected locale between search_dates() calls, so concurrent searches over text in different languages no longer return None or a date read in the wrong locale (#1371)
• Resolve the BST and HDT timezone abbreviations to the offsets the tz database gives them, UTC+1 (British Summer Time) and UTC-9 (Hawaii-Aleutian Daylight Time), instead of +11 and -9:30, which no zone goes by those names today; abbreviations that the tz database maps to more than one offset, such as CST and IST, keep their current offset. Text carrying these abbreviations keeps its wall clock but moves by 10 hours for BST and 30 minutes for HDT, which can put the parsed instant on a different day (#1366)
• Reject a %j (day of year) value that the parsed year does not have, instead of rolling it over into the next year, so "1999366" with date_formats=["%Y%j"] returns None rather than 2000-01-01. A format with no year directive is checked against the year strptime defaults to, 1900, which is not a leap year, so "366" with date_formats=["%j"] now returns None where it used to return January 1 (#1370)

Cleanups and internal improvements:

• Add CodSpeed benchmarks and a workflow that runs them, so a performance regression such as the quadratic backtracking fixed in 1.4.1 is reported on the pull request that introduces it (#1365)

1.4.2

New features

• Add an IGNORE_SURROUNDING_TEXT setting that, when enabled, retries parsing after ignoring leading and trailing words the language does not recognize, so a date wrapped in extra text such as "Published on 16/04/2019" is parsed (#1356)
• Add a strategy argument to search_dates() to choose the search strategy: the default "split" keeps the current behavior, while the new "ngram" strategy parses the longest sequences of tokens as dates for more predictable results on noisy text (#1351)

Fixes

• Do not read a two-digit number as a year once a later component has been found in year-first date orders, so the day in Japanese dates such as "4月20日" is no longer consumed as the year (#1358)
• Parse ISO 8601 dates (those starting with a four-digit year) in month-day order even when an explicit day-first language such as it or fr is given, without needing to set PREFER_LOCALE_DATE_ORDER to False (#1352)
• Honor PREFER_DATES_FROM and RELATIVE_BASE when parsing with custom date_formats that use a two-digit year (%y) (#1342)
• Honor the %j (day of year) directive in date_formats instead of overwriting the parsed month and day with the current date (#1345)

... (truncated)

Changelog

Sourced from dateparser's changelog.

1.4.3 (2026-09-03)

Fixes:

• Make parsing thread-safe: parsing from several threads no longer raises an intermittent KeyError from the shared language caches, and a DATE_ORDER or RELATIVE_BASE value meant for one parse no longer leaks into the settings that other parses read, where it could make them return a wrong date (#1346)
• Do not share the language detector and the detected locale between search_dates() calls, so concurrent searches over text in different languages no longer return None or a date read in the wrong locale (#1371)
• Resolve the BST and HDT timezone abbreviations to the offsets the tz database gives them, UTC+1 (British Summer Time) and UTC-9 (Hawaii-Aleutian Daylight Time), instead of +11 and -9:30, which no zone goes by those names today; abbreviations that the tz database maps to more than one offset, such as CST and IST, keep their current offset. Text carrying these abbreviations keeps its wall clock but moves by 10 hours for BST and 30 minutes for HDT, which can put the parsed instant on a different day (#1366)
• Reject a %j (day of year) value that the parsed year does not have, instead of rolling it over into the next year, so "1999366" with date_formats=["%Y%j"] returns None rather than 2000-01-01. A format with no year directive is checked against the year strptime defaults to, 1900, which is not a leap year, so "366" with date_formats=["%j"] now returns None where it used to return January 1 (#1370)

Cleanups and internal improvements:

• Add CodSpeed benchmarks and a workflow that runs them, so a performance regression such as the quadratic backtracking fixed in 1.4.1 is reported on the pull request that introduces it (#1365)

1.4.2 (2026-08-04)

New features:

• Add an IGNORE_SURROUNDING_TEXT setting that, when enabled, retries parsing after ignoring the leading and trailing words the language does not recognize, so a date wrapped in extra text such as "Published on 16/04/2019" is parsed (#1356)
• Add a strategy argument to search_dates() to choose the search strategy: the default "split" keeps the current behavior, while the new "ngram" strategy parses the longest sequences of tokens as dates for more predictable results on noisy text (#1351)

... (truncated)

Commits

• 9ce60b1 1.4.3 release (#1376)
• 27182cd Avoid shared language_detector and locale cache in search_dates (#1371)
• 6b5f17d Reject a %j day of year that the parsed year does not have (#271) (#1370)
• cd9bb13 Make thread-safe (#1346)
• de7b986 tests: add CodSpeed benchmarks (#1365)
• 764091d Fix timezone ambiguity using tz database (#1322) (#1366)
• b1f603e 1.4.2 release (#1361)
• ff3dcb9 Specify README header content as RST (#1360)
• d886e23 Parse dates wrapped in harmless extra text (Close #518) (#1356)
• 06ca032 Do not read the day as a two-digit year after the year position (#1358)
• Additional commits viewable in compare view

Updates eventsourcing from 9.5.2 to 9.5.6

Commits

• bf66595 Increaesd version to 9.9.6.
• 09861aa Merge pull request #313 from MRozmus/fix-psycopg-336-notifies
• 5cbf654 Fixed Postgres subscriptions hanging on exit with psycopg 3.3.6.
• 575d42c Updated pyproject.toml and poetry.lock to use packaged `eventsourcing-uma...
• be45e61 Added RUF063 to excluded linter rules for class annotations.
• 35d3da8 Synced dependencies to version 9.5.5 and added event comparison helper method...
• 524de24 Refactored singledispatchmethod to improve type handling, thread safety, an...
• acd1130 Replaced TypeVar from typing with typing_extensions, and updated depend...
• 20bc131 Drop Python 3.10, update dependencies.
• a0fa963 Added CPY001 to excluded linter rules in pyproject.toml.
• Additional commits viewable in compare view

Updates gliner2 from 2.0.0 to 2.1.0

Release notes

Sourced from gliner2's releases.

v2.1.0

GLiNER2 2.1.0 is on PyPI. Paste the note below into a Hugging Face model card or discussion.

# GLiNER2 2.1.0
Released 9 October 2026. Install with:
pip install -U "gliner2==2.1.0"
</code></pre>
<p>This is the library release after <a href="https://pypi.org/project/gliner2/2.0.0/">gliner2 2.0.0</a> (24 August 2026, the GLiNER2.5 release). It is built from <a href="https://github.com/fastino-ai/GLiNER2">fastino-ai/GLiNER2</a> <code>main</code> at <code>b86d35c</code>. Checkpoint weights are unchanged. What changed is inference, training, schemas, and packaging.</p>
<p>Full history: <a href="https://github.com/fastino-ai/GLiNER2/compare/v2.0.0...b86d35c">v2.0.0...b86d35c</a></p>
<h2>Highlights</h2>
<ul>
<li><strong>Transformers 5.</strong> <code>transformers&gt;=4.43,&lt;6</code> on every install profile. The <code>disentangled-flash</code> extra requires <code>transformers&gt;=4.57,&lt;6</code>.</li>
<li><strong>DisentangledFlash attention</strong> for DeBERTa encoders, including an automatic packed (unpadded) path and a differentiable training path.</li>
<li><strong>Faster span scoring and decoder synchronization.</strong></li>
<li><strong>Schemas from JSON now match the Python builder</strong>, including thresholds, <code>top_k</code>, label descriptions, prompts, and few-shot examples.</li>
<li><strong>Precomputed encoder states</strong> can be passed into every batch inference API, and a boundary model can be loaded without the encoder.</li>
<li>Training fixes for gold-label corruption, Indic tokenization, phantom entity spans, half-precision NaNs, record choice fields, and LoRA unload.</li>
</ul>
<h2>Faster inference</h2>
<h3>DisentangledFlash</h3>
<p>New load option <code>attention_backend</code>:</p>
<table>
<thead>
<tr>
<th>Value</th>
<th>Behavior</th>
</tr>
</thead>
<tbody>
<tr>
<td><code>&quot;standard&quot;</code></td>
<td>Transformers attention</td>
</tr>
<tr>
<td><code>&quot;flashdeberta&quot;</code></td>
<td>FlashDeBERTa. <code>use_flashdeberta</code> remains a compatibility alias</td>
</tr>
<tr>
<td><code>&quot;disentangled_flash&quot;</code></td>
<td><a href="https://github.com/delyan-boychev/disentangled-flash">DisentangledFlash</a></td>
</tr>
</tbody>
</table>
<pre lang="python"><code>from gliner2 import GLiNER2
model = GLiNER2.from_pretrained(
&quot;fastino/gliner2-base-v1&quot;,
attention_backend=&quot;disentangled_flash&quot;,
)
</code></pre>
<p>Install the extra with <code>pip install &quot;gliner2[disentangled-flash]&quot;</code>. On CUDA, FP16 or BF16 is what makes the flash kernels pay off.</p>
<p>Packed layout is on by default in auto mode. A batch uses the unpadded kernel when padding is at least 15% and the batch is large enough (<code>B * L2 &gt;= 221</code>, measured on an H200). Smaller batches stay on the padded path. Tune it with <code>disentangled_flash_packed</code>, <code>disentangled_flash_min_padding</code>, <code>disentangled_flash_min_work</code>, and <code>disentangled_flash_tuning</code>.</p>
<!-- raw HTML omitted -->
</blockquote>
<p>... (truncated)</p>
</details>
<details>
<summary>Commits</summary>
<ul>
<li><a href="https://github.com/fastino-ai/GLiNER2/commit/b86d35c02f2c1a4d9840ea50197bd1b6d13b031e&quot;&gt;&lt;code&gt;b86d35c&lt;/code&gt;&lt;/a> Merge pull request <a href="https://redirect.github.com/fastino-ai/GLiNER2/issues/189&quot;&gt;#189&lt;/a> from MattThomas-fastino/transformers-v5-support</li>
<li><a href="https://github.com/fastino-ai/GLiNER2/commit/ec0e31271d1a627e135e91ccc2932a97cbd68494&quot;&gt;&lt;code&gt;ec0e312&lt;/code&gt;&lt;/a> Support Transformers 5 and refresh dependency caps</li>
<li><a href="https://github.com/fastino-ai/GLiNER2/commit/08c8c8df0803075035ac6a02e8ee8c5554d2e06b&quot;&gt;&lt;code&gt;08c8c8d&lt;/code&gt;&lt;/a> Merge pull request <a href="https://redirect.github.com/fastino-ai/GLiNER2/issues/188&quot;&gt;#188&lt;/a> from delyan-boychev/optimize-decoder-sync</li>
<li><a href="https://github.com/fastino-ai/GLiNER2/commit/2763f94753ef910d77e71e4bcc3018207508c640&quot;&gt;&lt;code&gt;2763f94&lt;/code&gt;&lt;/a> Merge pull request <a href="https://redirect.github.com/fastino-ai/GLiNER2/issues/187&quot;&gt;#187&lt;/a> from kfastino/fix/cls-true-label-synthetic</li>
<li><a href="https://github.com/fastino-ai/GLiNER2/commit/467ea4fdbbb8ee3077b545333be2a23c2331d8a5&quot;&gt;&lt;code&gt;467ea4f&lt;/code&gt;&lt;/a> Add decoder synchronization benchmark matrix</li>
<li><a href="https://github.com/fastino-ai/GLiNER2/commit/f59fab04eb1d3c95fe223677e289820ea677c124&quot;&gt;&lt;code&gt;f59fab0&lt;/code&gt;&lt;/a> Merge remote-tracking branch 'upstream/main' into optimize-decoder-sync</li>
<li><a href="https://github.com/fastino-ai/GLiNER2/commit/0b67de781678825c8cd4fda5a9a25e5563ff3137&quot;&gt;&lt;code&gt;0b67de7&lt;/code&gt;&lt;/a> Merge pull request <a href="https://redirect.github.com/fastino-ai/GLiNER2/issues/143&quot;&gt;#143&lt;/a> from delyan-boychev/optimize-span-scoring</li>
<li><a href="https://github.com/fastino-ai/GLiNER2/commit/879fc2d4d3e9ef481dcc1368a0eebe462e0d802d&quot;&gt;&lt;code&gt;879fc2d&lt;/code&gt;&lt;/a> Optimize boundary decoder synchronization</li>
<li><a href="https://github.com/fastino-ai/GLiNER2/commit/cf348748194dd878831390d4bd674bfe651b1e9b&quot;&gt;&lt;code&gt;cf34874&lt;/code&gt;&lt;/a> Optimize accelerator decoder synchronization</li>
<li><a href="https://github.com/fastino-ai/GLiNER2/commit/0fa0e8870cfc30495b5c3870e0fe87dcef7af99a&quot;&gt;&lt;code&gt;0fa0e88&lt;/code&gt;&lt;/a> Merge remote-tracking branch 'origin/main' into optimize-span-scoring</li>
<li>Additional commits viewable in <a href="https://github.com/fastino-ai/GLiNER2/compare/v2.0.0...v2.1.0&quot;&gt;compare view</a></li>
</ul>
</details>
<br />
Updates peft from 0.20.0 to 0.21.2

Release notes
Sourced from peft's releases.

v0.21.2
This is a PEFT release fixes an issue that prevented encoder-decoder models to work when using Transformers ≥ 5.18.0.
Changes:

#3845

v0.21.1
This is a small PEFT release to enable Tensor Parallel (TP) to work properly with PEFT. It requires Transformers ≥ 5.17.0 to work.
Changes:

#3614
#3790

v0.21.0
Highlights

New methods
Riemannian-preconditional LoRA optimizer
Riemannian-preconditioned LoRA optimizer (#3382), contributed by @​smellslikeml, is a new LoRA-specific optimizer similar to LoRA-FA or LoRA+ that applies a Riemannian pre-conditioning on the gradients to improve feature learning and avoiding separate tuning of learning rates for LoRA's A and B matrices.
LoRA variant: Knowledge-Aware Singular-Value Adaption (KaSA)
KaSA aims at identifying knowledge aware singular values from the base weight. It then removes noisy knowledge from those weights and specifically trains the singular values most relevant to the task. Implemented in #3446 by @​robbiebusinessacc and @​iambogeumkim.
Super-Tuning (#3518)
Super-Tuning uses calibration-free magnitude scoring to allocate a sparse tuning budget. With BottomK selection, Super adapts low-magnitude weights, motivated by the idea that they provide degrees of freedom for learning a new task while leaving larger-magnitude pretrained weights untouched. Supra adds LoRA on top of this sparse update. It is a good fit for exploring sparsity alongside higher-rank adaptation, particularly on reasoning-heavy fine-tuning tasks. Thanks to @​smellslikeml.
ShadowPEFT (#3354)
ShadowPEFT runs a small shadow network alongside the frozen base model.  That shadow state gets updated at every decoder layer, so adaptation is shared across the stack instead of a separate low-rank update on each targeted linear layer (like LoRA). It’s a good choice when you care about downstream or OOD performance and want to stay in the same trainable-parameter range as LoRA, or when you want the small shadow on the edge and the frozen base in the cloud. This works best if the shadow model starts from a small, pretrained model. Contributed by @​SeanLee97.
Enhancements


Speed up MoE target parameter computation (#3577): For the common use case of one active adapter, weight and MoE LoRA delta computation is folded into a single baddbmm operation. Expect slight numerical differences in this case.


Support torchao v0.18.0 by supporting the newer Int8Tensor, Float8Tensor, etc. tensor subclasses.


Orthogonal Subspace Fine-tuning received a significant update in #3524 reducing the memory footprint by 22% and the training time by 46% according to the MetaMathQA benchmark while keeping task accuracy and forgetting stable. Expect slight numerical differences, especially in fp16, bf16.


huggingface/peft#3639: Thanks to @​amogh-nagri-11, prompt learning methods now allow to delete an adapter. For adaption prompt, adding or loading a second adapter no longer automatically activates it, in line with how other PEFT methods handle this.


Changes


... (truncated)


Commits

eaa35c7 Patch release v0.21.2 changes (#3855)
c9e7474 Patch release v0.21.1 changes (#3833)
9dc6fa2 Release: 0.21.0 (#3748)
20feaa7 FIX Loading scalar adapter state entries (#3699)
c9185a3 ENH Warn about unused rank / alpha pattern keys (#3588)
578142b DOC Fix multiple typos, wrong arguments and links (#3687)
6aa4944 Method comparison: Add UniLoRA benchmark to image-gen (#3740)
b3d183f Mixed model: Delete every auxiliary adapter and resolve the remaining active ...
e99fdd2 FIX Forward autocast_adapter_dtype on all PeftModel subclasses (#3668)
a5e4c58 ENH Reject MiSS and PSOFT invalid ranks (#3708)
Additional commits viewable in compare view



Updates fastapi from 0.142.2 to 0.161.1

Release notes
Sourced from fastapi's releases.

0.161.1
Fixes

🐛 Fix OpenAPI schemas for standalone nullable form unions. PR #16598 by @​tiangolo.

0.161.0
Features

✨ Add support for combining unions of form models. PR #16596 by @​tiangolo.

0.160.0
Query, header, cookie, and form parameter models now apply their extra configuration to all submitted fields in the same parameter location, consistent with Pydantic behavior. This includes fields declared in other models, individual parameters, or dependencies, which were previously excluded.
With extra="forbid", fields outside the model are rejected even if another parameter declares them. With extra="allow", those fields are included as model extras. With the default extra="ignore", they are ignored.
Breaking Changes

🐛 Apply each parameter model's extra configuration to sibling fields. PR #16594 by @​tiangolo.

Internal

⬆ Bump certifi from 2026.1.4 to 2026.10.19. PR #16571 by @​dependabot[bot].

0.159.1
Features

✨ Add support for optional unions of form models. PR #16586 by @​tiangolo.

0.159.0
Features

✨ Add support for optional query, header, and cookie parameter models. PR #16584 by @​tiangolo.

0.158.2
Features

✨ Add support for optional form models. PR #16582 by @​tiangolo.

0.158.1
Fixes

🐛 Preserve form media types in OpenAPI. PR #16579 by @​tiangolo.

0.158.0
Features

✨ Add support for combining form models. PR #16577 by @​tiangolo.

0.157.0
Features


... (truncated)


Commits

69ddc2b 🔖 Release version 0.161.1 (#16599)
0cafa07 📝 Update release notes
f90d036 🐛 Fix OpenAPI schemas for standalone nullable form unions (#16598)
80c8742 🔖 Release version 0.161.0 (#16597)
b105209 📝 Update release notes
e42ac0a ✨ Add support for combining unions of form models (#16596)
addd077 🔖 Release version 0.160.0 (#16595)
352a4ed 📝 Update release notes
91a335c 🐛 Apply each parameter model's extra configuration to sibling fields (#16594)
ce6ee83 📝 Update release notes
Additional commits viewable in compare view



Updates langchain from 1.4.5 to 1.4.7

Release notes
Sourced from langchain's releases.

langchain==1.4.7
Changes since langchain==1.4.6
chore(langchain): update lockfile (#41402)
release(langchain): 1.4.7 (#41401)
chore(deps): bump certifi from 2025.11.12 to 2026.10.19 in /libs/langchain_v1 (#41370)
feat(core,langchain,standard-tests): read tool-call-delta in stream consumers (#41246)
chore(deps): bump aiohttp from 3.14.4 to 3.14.5 in /libs/langchain_v1 (#41360)
chore(deps): bump aiohttp from 3.14.3 to 3.14.4 in /libs/langchain_v1 (#41330)
chore(deps): bump mcp from 2.1.1 to 2.2.0 in /libs/langchain_v1 (#41270)
feat(core): accept address inside LangChainTracer (#41006)
langchain==1.4.6
Changes since langchain==1.4.5
release(langchain): 1.4.6 (#41247)
feat(langchain): add opt-in per-call interrupts to HITL middleware (#41062)



Commits

fbb8a35 chore(langchain): update lockfile (#41402)
ad92e36 release(langchain): 1.4.7 (#41401)
bdf0f53 release(core): 1.6.13 (#41397)
59be9ee chore(text-splitters): bump certifi to 2026.10.19 (#41393)
82ef933 release(perplexity): 1.4.2 (#41392)
fc23e66 chore(model-profiles): refresh model profile data (#41354)
331e6be chore(deps): bump certifi from 2025.11.12 to 2026.10.19 in /libs/partners/ant...
144b0b4 chore(deps): bump certifi from 2025.8.3 to 2026.10.19 in /libs/partners/chrom...
9b5ae66 chore(deps): bump certifi from 2025.8.3 to 2026.10.19 in /libs/partners/deeps...
c773a1b chore(deps): bump certifi from 2025.8.3 to 2026.10.19 in /libs/partners/exa (...
Additional commits viewable in compare view



Updates langchain-community from 0.4.1 to 0.4.2

Release notes
Sourced from langchain-community's releases.

langchain-community==0.4.2
Sunsetting langchain-community
langchain-community is being sunset. See langchain-ai/langchain-community#674 for details and guidance. Thank you to everyone who has contributed integrations, fixes, reviews, and maintenance over the years.
What's Changed

chore: modernize infra by @​mdrxy in langchain-ai/langchain-community#402
Add PMID from PubMedAPIWrapper utility by @​davale88 in langchain-ai/langchain-community#43
gitlab api wrapper[patch]: fix create_file eats the first letter by @​martra in langchain-ai/langchain-community#153
Update lancedb.py by @​AmirBraham in langchain-ai/langchain-community#108
Correcting OpenSearchVectorStore documentation for AsyncClient import problem by @​georg-wind in langchain-ai/langchain-community#252
Update test_duckduckdgo_search_api.py import by @​theprashasst in langchain-ai/langchain-community#248
Fixed typos. by @​oumizx in langchain-ai/langchain-community#277
Correct misleading CosmosDB connection string validation error message by @​lasagna0 in langchain-ai/langchain-community#315
fix: fix score threshold is zero scenario by @​MichaelLi65535 in langchain-ai/langchain-community#398
Fix broken link of partition docs of unstructured package by @​mhihasan in langchain-ai/langchain-community#354
fix(loader): fix TextLoader doc string indentation by @​vneogi199 in langchain-ai/langchain-community#360
Fixed Syntax Error in Trino dialect by @​saathwik-tk in langchain-ai/langchain-community#399
fix: message line regex by @​martinschaer in langchain-ai/langchain-community#196
community: remove integration as the product has changed by @​Xe in langchain-ai/langchain-community#159
Update sql_database.py to support materialized view by @​groundrace in langchain-ai/langchain-community#281
RSS Feed Document Loader Publish Date Fallback by @​mariomeyer in langchain-ai/langchain-community#219
fix: lint by @​mdrxy in langchain-ai/langchain-community#405
chore(infra): add auto-labelers by @​mdrxy in langchain-ai/langchain-community#406
feat: add cost by 1k tokens for sonnet 4.5 by @​higorgardin in langchain-ai/langchain-community#391
fix: Refactor query_builder to use limit method by @​userNanni in langchain-ai/langchain-community#377
feat: migrate document_loaders from rapidocr_onnxruntime to rapidocr by @​athoik in langchain-ai/langchain-community#375
Deprecate Oracle AI Vector Search classes by @​fileames in langchain-ai/langchain-community#283
Fix: Add default factory to api_wrapper of GoogleSerperRun by @​duccm94 in langchain-ai/langchain-community#157
community[fix]: support spark x1 reasoning model's 'reasoning_content' field. by @​vsxd in langchain-ai/langchain-community#251
fix: use uuid1() instead of uuid4() for Cassandra TIMEUUID columns by @​woohhan in langchain-ai/langchain-community#225
Import chat_models from langchain_community in langchain-ai/langchain-community#198
fix: Retrieve all Slack channels visible to the client by @​m1kl0sh in langchain-ai/langchain-community#146
feat: fix elevenlabs >=2.3.0 api change by @​ebwood in langchain-ai/langchain-community#102
fix: write image data to buffer before checking for empty BytesIO in image extraction by @​mighoocool in langchain-ai/langchain-community#193
align SupabaseVectorStore ID handling with JS lib by @​taishikato in langchain-ai/langchain-community#137
fix(document_loaders): yt api as per latest version by @​lokeshparab in langchain-ai/langchain-community#206
feat: Add comprehensive routing support to OpenSearchVectorSearch by @​yukiharada1228 in langchain-ai/langchain-community#349
fix(chat_models): parallel tool calls in ChatTongyi when incremental streaming output is enabled. by @​TBice123123 in langchain-ai/langchain-community#111
fix(infra): remove unused dataclasses-json dependency by @​noirbee in langchain-ai/langchain-community#270
chore: update README.md by @​mdrxy in langchain-ai/langchain-community#408
feat(agent_toolkits): Make WebClient injectable for Slack toolkit by @​m1kl0sh in langchain-ai/langchain-community#145
feat(document_loaders): add flexible timeout to PlaywrightURLLoader by @​parthsompura in langchain-ai/langchain-community#104
feat: add SSL verification option to recursive link fetching by @​rajasblack in