> [!WARNING]
> **BREAKING:** Simplifies message normalization to single consistent
path, requiring partner package updates
**Key Changes:**
- Consistent multimodal handling:
- OpenAI `image_url` blocks pass through unchanged (broad compatibility)
- OpenAI `input_audio` and `file` blocks convert to v1 standard
equivalents
- Legacy v0 multimodal blocks convert to v1 standard
- Everything else passes through unchanged
- Partner packages must update content block parsing logic
**Partner Updates**
`output_version` affects how messages are serialized into `.content`.
`_normalize_messages()` will now upgrade v0 content to v1, so, all
partners now receive v1 format input regardless of `output_version`.
Migration:
- Partner packages must update to handle v1 input content blocks
- `output_version` still controls serialization format of responses
(unchanged)
---------
Co-authored-by: Chester Curme <chester.curme@gmail.com>
Extensible registry system for translating AI message content blocks
from various model providers. Refactors the way provider-specific
content is handled, moving from hardcoded logic to a plugin-like
architecture.
Having vercel attempt to deploy on each commit (even if unrelated to
docs) was getting annoying. Options:
- `[skip-preview]`
- `[no-preview]`
- `[skip-deploy]`
Full example: `fix(core): resolve memory leak [no-preview]`
Put in `core.utils` this time to prevent other circular import issues
present in the `normalize()` rfc:
`base` imports `content`
`content` imports `ensure_id()` from `base`
re: #32589 cc: @ccurme
- Rename namespace: `messages.content_blocks` -> `messages.content`
- Prefixes and ID logic are now in `messages.common` instead of
`AIMessage` since the logic is shared between messages and message
content. Did this instead of `utils` due to circular import problems
that were hairy
* Create usage metadata on
[`message_delta`](https://docs.anthropic.com/en/docs/build-with-claude/streaming#event-types)
instead of at the beginning. Consequently, token counts are not included
during streaming but instead at the end. This allows for accurate
reporting of server-side tool usage (important for billing)
* Add some clarifying comments
* Fix some outstanding Pylance warnings
* Remove unnecessary `text` popping in thinking blocks
* Also now correctly reports `input_cache_read`/`input_cache_creation`
as a result
When citations are returned from streaming, they include a `file_id:
null` field in their `content_block_location` structure.
When these citations are passed back to the API in subsequent messages,
the API rejects them with "Extra inputs are not permitted" for the
`file_id` field.
**Description:**
Corrected LangGraph documentation link (changed to “guides”), and added
a link to LangGraph JS how-to guides for clarity.
**Issue:**
N/A
**Dependencies:**
None
---------
Co-authored-by: Mason Daugherty <mason@langchain.dev>
The appropriate `ToolNode` attribute for error handling is called
`handle_tool_errors` instead of `handle_tool_error`.
For further info see [ToolNode source code in
LangGraph](https://github.com/langchain-ai/langgraph/blob/main/libs/prebuilt/langgraph/prebuilt/tool_node.py#L255)
**Twitter handle:** gitaroktato
- [x] **Add tests and docs**: If you're adding a new integration, you
must include:
1. A test for the integration, preferably unit tests that do not rely on
network access,
2. An example notebook showing its use. It lives in
`docs/docs/integrations` directory.
- [x] **Lint and test**: Run `make format`, `make lint` and `make test`
from the root of the package(s) you've modified. **We will not consider
a PR unless these three are passing in CI.** See [contribution
guidelines](https://python.langchain.com/docs/contributing/) for more.
Additional guidelines:
- Make sure optional dependencies are imported within a function.
- Please do not add dependencies to `pyproject.toml` files (even
optional ones) unless they are **required** for unit tests.
- Most PRs should not touch more than one package.
- Changes should be backwards compatible.
## Description
This PR adds support for custom header patterns in
`MarkdownHeaderTextSplitter`, allowing users to define non-standard
Markdown header formats (like `**Header**`) and specify their hierarchy
levels.
**Issue:** Fixes#22738
**Dependencies:** None - this change has no new dependencies
**Key Changes:**
- Added optional `custom_header_patterns` parameter to support
non-standard header formats
- Enable splitting on patterns like `**Header**` and `***Header***`
- Maintain full backward compatibility with existing usage
- Added comprehensive tests for custom and mixed header scenarios
## Example Usage
```python
from langchain_text_splitters import MarkdownHeaderTextSplitter
headers_to_split_on = [
("**", "Chapter"),
("***", "Section"),
]
custom_header_patterns = {
"**": 1, # Level 1 headers
"***": 2, # Level 2 headers
}
splitter = MarkdownHeaderTextSplitter(
headers_to_split_on=headers_to_split_on,
custom_header_patterns=custom_header_patterns,
)
# Now **Chapter 1** is treated as a level 1 header
# And ***Section 1.1*** is treated as a level 2 header
```
## Testing
- ✅ Added unit tests for custom header patterns
- ✅ Added tests for mixed standard and custom headers
- ✅ All existing tests pass (backward compatibility maintained)
- ✅ Linting and formatting checks pass
---
The implementation provides a flexible solution while maintaining the
simplicity of the existing API. Users can continue using the splitter
exactly as before, with the new functionality being entirely opt-in
through the `custom_header_patterns` parameter.
---------
Co-authored-by: Mason Daugherty <mason@langchain.dev>
Co-authored-by: Claude <noreply@anthropic.com>
Supersedes #32461
Fixed incorrect input token reporting during streaming when tools are
used. Previously, input tokens were counted at `message_start` before
tool execution, leading to inaccurate counts. Now input tokens are
properly deferred until `message_delta` (completion), aligning with
Anthropic's billing model and SDK expectations.
**Before Fix:**
- Streaming with tools: Input tokens = 0 ❌
- Non-streaming with tools: Input tokens = 472 ✅
**After Fix:**
- Streaming with tools: Input tokens = 472 ✅
- Non-streaming with tools: Input tokens = 472 ✅
Aligns with Anthropic's SDK expectations. The SDK handles input token
updates in `message_delta` events:
```python
# https://github.com/anthropics/anthropic-sdk-python/blob/main/src/anthropic/lib/streaming/_messages.py
if event.usage.input_tokens is not None:
current_snapshot.usage.input_tokens = event.usage.input_tokens
```
