github/langchain
1
0
Fork 0
mirror of https://github.com/hwchase17/langchain.git synced 2026-06-09 10:17:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

feat(anthropic): delegate cache_control kwarg to anthropic top-level param (#35967)

Browse Source
This commit is contained in:
ccurme
2026-03-17 10:49:03 -04:00
committed by GitHub
parent 5a2c999855
commit 55711b010b
5 changed files with 11 additions and 362 deletions
Download Patch File Download Diff File Expand all files Collapse all files

97
libs/partners/anthropic/langchain_anthropic/chat_models.py
View File

@@ -688,65 +688,6 @@ def _format_messages(
return system, formatted_messages
def _collect_code_execution_tool_ids(formatted_messages: list[dict]) -> set[str]:
"""Collect tool_use IDs that were called by code_execution.
These blocks cannot have cache_control applied per Anthropic API requirements.
"""
code_execution_tool_ids: set[str] = set()
for message in formatted_messages:
if message.get("role") != "assistant":
continue
content = message.get("content", [])
if not isinstance(content, list):
continue
for block in content:
if not isinstance(block, dict):
continue
if block.get("type") != "tool_use":
continue
caller = block.get("caller")
if isinstance(caller, dict):
caller_type = caller.get("type", "")
if caller_type.startswith("code_execution"):
tool_id = block.get("id")
if tool_id:
code_execution_tool_ids.add(tool_id)
return code_execution_tool_ids
def _is_code_execution_related_block(
block: dict,
code_execution_tool_ids: set[str],
) -> bool:
"""Check if a content block is related to code_execution.
Returns True for blocks that should NOT have cache_control applied.
"""
if not isinstance(block, dict):
return False
block_type = block.get("type")
# tool_use blocks called by code_execution
if block_type == "tool_use":
caller = block.get("caller")
if isinstance(caller, dict):
caller_type = caller.get("type", "")
if caller_type.startswith("code_execution"):
return True
# tool_result blocks for code_execution called tools
if block_type == "tool_result":
tool_use_id = block.get("tool_use_id")
if tool_use_id and tool_use_id in code_execution_tool_ids:
return True
return False
class AnthropicContextOverflowError(anthropic.BadRequestError, ContextOverflowError):
"""BadRequestError raised when input exceeds Anthropic's context limit."""
@@ -1127,44 +1068,6 @@ class ChatAnthropic(BaseChatModel):
system, formatted_messages = _format_messages(messages)
# If cache_control is provided in kwargs, add it to the last eligible message
# block (Anthropic requires cache_control to be nested within a message block).
# Skip blocks related to code_execution as they cannot have cache_control.
cache_control = kwargs.pop("cache_control", None)
if cache_control and formatted_messages:
# Collect tool IDs called by code_execution
code_execution_tool_ids = _collect_code_execution_tool_ids(
formatted_messages
)
cache_applied = False
for formatted_message in reversed(formatted_messages):
if cache_applied:
break
content = formatted_message.get("content")
if isinstance(content, list) and content:
# Find last eligible block (not code_execution related)
for block in reversed(content):
if isinstance(block, dict):
if _is_code_execution_related_block(
block, code_execution_tool_ids
):
continue
block["cache_control"] = cache_control
cache_applied = True
break
elif isinstance(content, str):
formatted_message["content"] = [
{
"type": "text",
"text": content,
"cache_control": cache_control,
}
]
cache_applied = True
# If we didn't find an eligible block we silently drop the control.
# Anthropic would reject a payload with cache_control on
# code_execution blocks.
payload = {
"model": self.model,
"max_tokens": self.max_tokens,