Fixes a streaming bug where models like Qwen3 (using OpenAI interface)
send tool call chunks with inconsistent indices, resulting in
duplicate/erroneous tool calls instead of a single merged tool call.
## Problem
When Qwen3 streams tool calls, it sends chunks with inconsistent `index`
values:
- First chunk: `index=1` with tool name and partial arguments
- Subsequent chunks: `index=0` with `name=None`, `id=None` and argument
continuation
The existing `merge_lists` function only merges chunks when their
`index` values match exactly, causing these logically related chunks to
remain separate, resulting in multiple incomplete tool calls instead of
one complete tool call.
```python
# Before fix: Results in 1 valid + 1 invalid tool call
chunk1 = AIMessageChunk(tool_call_chunks=[
{"name": "search", "args": '{"query":', "id": "call_123", "index": 1}
])
chunk2 = AIMessageChunk(tool_call_chunks=[
{"name": None, "args": ' "test"}', "id": None, "index": 0}
])
merged = chunk1 + chunk2 # Creates 2 separate tool calls
# After fix: Results in 1 complete tool call
merged = chunk1 + chunk2 # Creates 1 merged tool call: search({"query": "test"})
```
## Solution
Enhanced the `merge_lists` function in `langchain_core/utils/_merge.py`
with intelligent tool call chunk merging:
1. **Preserves existing behavior**: Same-index chunks still merge as
before
2. **Adds special handling**: Tool call chunks with
`name=None`/`id=None` that don't match any existing index are now merged
with the most recent complete tool call chunk
3. **Maintains backward compatibility**: All existing functionality
works unchanged
4. **Targeted fix**: Only affects tool call chunks, doesn't change
behavior for other list items
The fix specifically handles the pattern where:
- A continuation chunk has `name=None` and `id=None` (indicating it's
part of an ongoing tool call)
- No matching index is found in existing chunks
- There exists a recent tool call chunk with a valid name or ID to merge
with
## Testing
Added comprehensive test coverage including:
- ✅ Qwen3-style chunks with different indices now merge correctly
- ✅ Existing same-index behavior preserved
- ✅ Multiple distinct tool calls remain separate
- ✅ Edge cases handled (empty chunks, orphaned continuations)
- ✅ Backward compatibility maintained
Fixes#31511.
<!-- START COPILOT CODING AGENT TIPS -->
---
💬 Share your feedback on Copilot coding agent for the chance to win a
$200 gift card! Click
[here](https://survey.alchemer.com/s3/8343779/Copilot-Coding-agent) to
start the survey.
---------
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: mdrxy <61371264+mdrxy@users.noreply.github.com>
Co-authored-by: Mason Daugherty <github@mdrxy.com>
Co-authored-by: Mason Daugherty <mason@langchain.dev>
## Problem
ChatLiteLLM encounters a `ValidationError` when using cache on
subsequent calls, causing the following error:
```
ValidationError(model='ChatResult', errors=[{'loc': ('generations', 0, 'type'), 'msg': "unexpected value; permitted: 'ChatGeneration'", 'type': 'value_error.const', 'ctx': {'given': 'Generation', 'permitted': ('ChatGeneration',)}}])
```
This occurs because:
1. The cache stores `Generation` objects (with `type="Generation"`)
2. But `ChatResult` expects `ChatGeneration` objects (with
`type="ChatGeneration"` and a required `message` field)
3. When cached values are retrieved, validation fails due to the type
mismatch
## Solution
Added graceful handling in both sync (`_generate_with_cache`) and async
(`_agenerate_with_cache`) cache methods to:
1. **Detect** when cached values contain `Generation` objects instead of
expected `ChatGeneration` objects
2. **Convert** them to `ChatGeneration` objects by wrapping the text
content in an `AIMessage`
3. **Preserve** all original metadata (`generation_info`)
4. **Allow** `ChatResult` creation to succeed without validation errors
## Example
```python
# Before: This would fail with ValidationError
from langchain_community.chat_models import ChatLiteLLM
from langchain_community.cache import SQLiteCache
from langchain.globals import set_llm_cache
set_llm_cache(SQLiteCache(database_path="cache.db"))
llm = ChatLiteLLM(model_name="openai/gpt-4o", cache=True, temperature=0)
print(llm.predict("test")) # Works fine (cache empty)
print(llm.predict("test")) # Now works instead of ValidationError
# After: Seamlessly handles both Generation and ChatGeneration objects
```
## Changes
- **`libs/core/langchain_core/language_models/chat_models.py`**:
- Added `Generation` import from `langchain_core.outputs`
- Enhanced cache retrieval logic in `_generate_with_cache` and
`_agenerate_with_cache` methods
- Added conversion from `Generation` to `ChatGeneration` objects when
needed
-
**`libs/core/tests/unit_tests/language_models/chat_models/test_cache.py`**:
- Added test case to validate the conversion logic handles mixed object
types
## Impact
- **Backward Compatible**: Existing code continues to work unchanged
- **Minimal Change**: Only affects cache retrieval path, no API changes
- **Robust**: Handles both legacy cached `Generation` objects and new
`ChatGeneration` objects
- **Preserves Data**: All original content and metadata is maintained
during conversion
Fixes#22389.
<!-- START COPILOT CODING AGENT TIPS -->
---
💡 You can make Copilot smarter by setting up custom instructions,
customizing its development environment and configuring Model Context
Protocol (MCP) servers. Learn more [Copilot coding agent
tips](https://gh.io/copilot-coding-agent-tips) in the docs.
---------
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: mdrxy <61371264+mdrxy@users.noreply.github.com>
Co-authored-by: Mason Daugherty <github@mdrxy.com>
Co-authored-by: Mason Daugherty <mason@langchain.dev>
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
**Description:** Fixes incorrect `num_skipped` count in the LangChain
indexing API. The current implementation only counts documents that
already exist in RecordManager (cross-batch duplicates) but fails to
count documents removed during within-batch deduplication via
`_deduplicate_in_order()`.
This PR adds tracking of the original batch size before deduplication
and includes the difference in `num_skipped`, ensuring that `num_added +
num_skipped` equals the total number of input documents.
**Issue:** Fixes incorrect document count reporting in indexing
statistics
**Dependencies:** None
Fixes#32272
---------
Co-authored-by: Alex Feel <afilippov@spotware.com>
Ensures proper reStructuredText formatting by adding the required blank
line before closing docstring quotes, which resolves the "Block quote
ends without a blank line; unexpected unindent" warning.
This PR fixes the PostgreSQL NUL byte issue that causes
`psycopg.DataError` when inserting documents containing `\x00` bytes
into PostgreSQL-based vector stores.
## Problem
PostgreSQL text fields cannot contain NUL (0x00) bytes. When documents
with such characters are processed by PGVector or langchain-postgres
implementations, they fail with:
```
(psycopg.DataError) PostgreSQL text fields cannot contain NUL (0x00) bytes
```
This commonly occurs when processing PDFs, documents from various
loaders, or text extracted by libraries like unstructured that may
contain embedded NUL bytes.
## Solution
Added `sanitize_for_postgres()` utility function to
`langchain_core.utils.strings` that removes or replaces NUL bytes from
text content.
### Key Features
- **Simple API**: `sanitize_for_postgres(text, replacement="")`
- **Configurable**: Replace NUL bytes with empty string (default) or
space for readability
- **Comprehensive**: Handles all problematic examples from the original
issue
- **Well-tested**: Complete unit tests with real-world examples
- **Backward compatible**: No breaking changes, purely additive
### Usage Example
```python
from langchain_core.utils import sanitize_for_postgres
from langchain_core.documents import Document
# Before: This would fail with DataError
problematic_content = "Getting\x00Started with embeddings"
# After: Clean the content before database insertion
clean_content = sanitize_for_postgres(problematic_content)
# Result: "GettingStarted with embeddings"
# Or preserve readability with spaces
readable_content = sanitize_for_postgres(problematic_content, " ")
# Result: "Getting Started with embeddings"
# Use in Document processing
doc = Document(page_content=clean_content, metadata={...})
```
### Integration Pattern
PostgreSQL vector store implementations should sanitize content before
insertion:
```python
def add_documents(self, documents: List[Document]) -> List[str]:
# Sanitize documents before insertion
sanitized_docs = []
for doc in documents:
sanitized_content = sanitize_for_postgres(doc.page_content, " ")
sanitized_doc = Document(
page_content=sanitized_content,
metadata=doc.metadata,
id=doc.id
)
sanitized_docs.append(sanitized_doc)
return self._insert_documents_to_db(sanitized_docs)
```
## Changes Made
- Added `sanitize_for_postgres()` function in
`langchain_core/utils/strings.py`
- Updated `langchain_core/utils/__init__.py` to export the new function
- Added comprehensive unit tests in
`tests/unit_tests/utils/test_strings.py`
- Validated against all examples from the original issue report
## Testing
All tests pass, including:
- Basic NUL byte removal and replacement
- Multiple consecutive NUL bytes
- Empty string handling
- Real examples from the GitHub issue
- Backward compatibility with existing string utilities
This utility enables PostgreSQL integrations in both langchain-community
and langchain-postgres packages to handle documents with NUL bytes
reliably.
Fixes#26033.
<!-- START COPILOT CODING AGENT TIPS -->
---
💬 Share your feedback on Copilot coding agent for the chance to win a
$200 gift card! Click
[here](https://survey.alchemer.com/s3/8343779/Copilot-Coding-agent) to
start the survey.
---------
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: mdrxy <61371264+mdrxy@users.noreply.github.com>
Co-authored-by: Mason Daugherty <github@mdrxy.com>
* **Description:** Updated `parse_result` logic to handle cases where
`self.first_tool_only` is `True` and multiple matching keys share the
same function name. Instead of returning the first match prematurely,
the method now prioritizes filtering results by the specified key to
ensure correct selection.
* **Issue:** #32100
---------
Co-authored-by: Mason Daugherty <github@mdrxy.com>
Co-authored-by: Mason Daugherty <mason@langchain.dev>
Previously, we hit an index out of range error with empty variable names
(accessing tag[0]), now we through a slightly nicer error
---------
Co-authored-by: Mason Daugherty <mason@langchain.dev>
Fixes#32042
## Summary
Fixes a critical bug in JSON Schema reference resolution that prevented
correctly dereferencing numeric components in JSON pointer paths,
specifically for list indices in `anyOf`, `oneOf`, and `allOf` arrays.
## Changes
- Fixed `_retrieve_ref` function in
`libs/core/langchain_core/utils/json_schema.py` to properly handle
numeric components
- Added comprehensive test function `test_dereference_refs_list_index()`
in `libs/core/tests/unit_tests/utils/test_json_schema.py`
- Resolved line length formatting issues
- Improved type checking and index validation for list and dictionary
references
## Key Improvements
- Correctly handles list index references in JSON pointer paths
- Maintains backward compatibility with existing dictionary numeric key
functionality
- Adds robust error handling for out-of-bounds and invalid indices
- Passes all test cases covering various reference scenarios
## Test Coverage
- Verified fix for `#/properties/payload/anyOf/1/properties/startDate`
reference
- Tested edge cases including out-of-bounds and negative indices
- Ensured no regression in existing reference resolution functionality
Resolves the reported issue with JSON Schema reference dereferencing for
list indices.
---------
Co-authored-by: open-swe-dev[bot] <open-swe-dev@users.noreply.github.com>
Co-authored-by: Mason Daugherty <github@mdrxy.com>
Co-authored-by: Mason Daugherty <mason@langchain.dev>
- **Description:** Ensure that the tool description is an empty string
when creating a Structured Tool from a Pydantic class in case no
description is provided
- **Issue:** Fixes#31606
---------
Co-authored-by: Mason Daugherty <mason@langchain.dev>
## Description
Currently when deserializing objects that contain non-deserializable
values, we throw an error. However, there are cases (e.g. proxies that
return response fields containing extra fields like Python datetimes),
where these values are not important and we just want to drop them.
Twitter handle: @hacubu
---------
Co-authored-by: Mason Daugherty <github@mdrxy.com>
**Description:**
Added an explicit validation step in
`langchain_core.vectorstores.utils._cosine_similarity` to raise a
`ValueError` if the input query or any embedding contains `NaN` values.
This prevents silent failures or unstable behavior during similarity
calculations, especially when using maximal_marginal_relevance.
**Issue**:
Fixes#31806
**Dependencies:**
None
---------
Co-authored-by: Azhagammal S C <azhagammal@kofluence.com>
Co-authored-by: Mason Daugherty <mason@langchain.dev>
Add additional hashing options to the indexing API, warn on SHA-1
Requires:
- Bumping langchain-core version
- bumping min langchain-core in langchain
---------
Co-authored-by: ccurme <chester.curme@gmail.com>
`Runnable`'s `Input` is contravariant so we need to enumerate all
possible inputs and it's not possible to put them in a `Union`.
Also, it's better to only require a runnable that
accepts`list[BaseMessage]` instead of a broader `Sequence[BaseMessage]`
as internally the runnable is only called with a list.
* Simplified Pydantic handling since Pydantic v1 is not supported
anymore.
* Replace use of deprecated v1 methods by corresponding v2 methods.
* Remove use of other deprecated methods.
* Activate mypy errors on deprecated methods use.
---------
Co-authored-by: Eugene Yurtsev <eyurtsev@gmail.com>
- **Description:**
- In _infer_arg_descriptions, the annotations dictionary contains string
representations of types instead of actual typing objects. This causes
_is_annotated_type to fail, preventing the correct description from
being generated.
- This is a simple fix using the get_type_hints method, which resolves
the annotations properly and is supported across all Python versions.
- **Issue:** #31051
---------
Co-authored-by: Eugene Yurtsev <eyurtsev@gmail.com>
https://github.com/langchain-ai/langchain/pull/31286 included an update
to the return type for `BaseChatModel.(a)stream`, from
`Iterator[BaseMessageChunk]` to `Iterator[BaseMessage]`.
This change is correct, because when streaming is disabled, the stream
methods return an iterator of `BaseMessage`, and the inheritance is such
that an `BaseMessage` is not a `BaseMessageChunk` (but the reverse is
true).
However, LangChain includes a pattern throughout its docs of [summing
BaseMessageChunks](https://python.langchain.com/docs/how_to/streaming/#llms-and-chat-models)
to accumulate a chat model stream. This pattern is implemented in tests
for most integration packages and appears in application code. So
https://github.com/langchain-ai/langchain/pull/31286 introduces mypy
errors throughout the ecosystem (or maybe more accurately, it reveals
that this pattern does not account for use of the `.stream` method when
streaming is disabled).
Here we revert just the change to the stream return type to unblock
things. A fix for this should address docs + integration packages (or if
we elect to just force people to update code, be explicit about that).
* It is possible to chain a `Runnable` with an `AsyncIterator` as seen
in `test_runnable.py`.
* Iterator and AsyncIterator Input/Output of Callables must be put
before `Callable[[Other], Any]` otherwise the pattern matching picks the
latter.
**Issue:**[
#309070](https://github.com/langchain-ai/langchain/issues/30970)
**Cause**
Arg type in python code
```
arg: Union[SubSchema1, SubSchema2]
```
is translated to `anyOf` in **json schema**
```
"anyOf" : [{sub schema 1 ...}, {sub schema 1 ...}]
```
The value of anyOf is a list sub schemas.
The bug is caused since the sub schemas inside `anyOf` list is not taken
care of.
The location where the issue happens is `convert_to_openai_function`
function -> `_recursive_set_additional_properties_false` function, that
recursively adds `"additionalProperties": false` to json schema which is
[required by OpenAI's strict function
calling](https://platform.openai.com/docs/guides/structured-outputs?api-mode=responses#additionalproperties-false-must-always-be-set-in-objects)
**Solution:**
This PR fixes this issue by iterating each sub schema inside `anyOf`
list.
A unit test is added.
**Twitter handle:** shengboma
If no one reviews your PR within a few days, please @-mention one of
baskaryan, eyurtsev, ccurme, vbarda, hwchase17.
---------
Co-authored-by: ccurme <chester.curme@gmail.com>
**Description:** Before this commit, if one record is batched in more
than 32k rows for sqlite3 >= 3.32 or more than 999 rows for sqlite3 <
3.31, the `record_manager.delete_keys()` will fail, as we are creating a
query with too many variables.
This commit ensures that we are batching the delete operation leveraging
the `cleanup_batch_size` as it is already done for `full` cleanup.
Added unit tests for incremental mode as well on different deleting
batch size.
**Description**: The 'inspect' package in python skips over the aliases
set in the schema of a pydantic model. This is a workound to include the
aliases from the original input.
**issue**: #31035
Cc: @ccurme @eyurtsev
---------
Co-authored-by: Chester Curme <chester.curme@gmail.com>
When aggregating AIMessageChunks in a stream, core prefers the leftmost
non-null ID. This is problematic because:
- Core assigns IDs when they are null to `f"run-{run_manager.run_id}"`
- The desired meaningful ID might not be available until midway through
the stream, as is the case for the OpenAI Responses API.
For the OpenAI Responses API, we assign message IDs to the top-level
`AIMessage.id`. This works in `.(a)invoke`, but during `.(a)stream` the
IDs get overwritten by the defaults assigned in langchain-core. These
IDs
[must](https://community.openai.com/t/how-to-solve-badrequesterror-400-item-rs-of-type-reasoning-was-provided-without-its-required-following-item-error-in-responses-api/1151686/9)
be available on the AIMessage object to support passing reasoning items
back to the API (e.g., if not using OpenAI's `previous_response_id`
feature). We could add them elsewhere, but seeing as we've already made
the decision to store them in `.id` during `.(a)invoke`, addressing the
issue in core lets us fix the problem with no interface changes.
Chat models currently implement support for:
- images in OpenAI Chat Completions format
- other multimodal types (e.g., PDF and audio) in a cross-provider
[standard
format](https://python.langchain.com/docs/how_to/multimodal_inputs/)
Here we update core to extend support to PDF and audio input in Chat
Completions format. **If an OAI-format PDF or audio content block is
passed into any chat model, it will be transformed to the LangChain
standard format**. We assume that any chat model supporting OAI-format
PDF or audio has implemented support for the standard format.
