github/langchain
1
0
Fork 0
mirror of https://github.com/hwchase17/langchain.git synced 2026-07-12 11:21:37 +00:00
Code Issues Packages Projects Releases Wiki Activity

docs(core): document positional-fallback fragility in bridge

Browse Source 
When a provider emits content_blocks without an `index` field (e.g.
anthropic `_stream` with coerce_content_to_string=True for pure text),
the bridge's positional-0 fallback merges successive chunks into one
block. This works correctly today because the only coerced-string path
in the anthropic integration is mutually exclusive with any structured
(indexed) emission. Document the scenario that would surface the gap
and the two ways to close it (native _stream_chat_model_events hook or
'continue on anonymous key' bridge rule) so a future integration
doesn't discover the edge case the hard way.
This commit is contained in:
Nick Hollon
2026-04-21 13:25:52 -04:00
parent 810def4fc5
commit 07c71ae632
1 changed files with 19 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
libs/core/langchain_core/language_models/_compat_bridge.py
View File

@@ -121,6 +121,25 @@ def _iter_protocol_blocks(msg: BaseMessage) -> list[tuple[Any, CompatBlock]]:
For finalized :class:`AIMessage`, also surfaces `invalid_tool_calls`
— which `AIMessage.content_blocks` currently omits from its return
value even though they are a defined protocol block type.
The positional fallback is a known fragility: when a provider emits
blocks without an `index` field (e.g. Anthropic's `_stream` with
`coerce_content_to_string=True`, where text chunks lose their
source-side index), every such chunk gets positional key 0 and
successive chunks merge into one block. This works correctly for
single-type streams (pure-text responses merge cleanly) because all
chunks share the same key and the open-block logic collapses them.
It would miscategorise a stream that mixed indexed structured
blocks with non-indexed coerced-text blocks, since an indexed
block with `index == 0` would collide with the anonymous text
block's positional-0 key. In the anthropic integration this
cannot currently occur: coerce-to-string mode is only selected
when no tools, thinking, or documents are present, and any of
those flips the stream to structured mode where every block
carries an integer index. A native `_stream_chat_model_events`
hook per provider (or a bridge-level "continue the open block when
the source has no identity" rule) would close the gap if another
integration ever emits mixed content.
"""
try:
raw = msg.content_blocks