PR #35813: fix(openai): skip malformed tool call chunks instead of dropping entire batch

• Repository: langchain-ai/langchain
• Author: gitbalaji
• State: closed | merged: False
• Link: https://github.com/langchain-ai/langchain/pull/35813

Description (problem / solution / changelog)

Summary

• Fixes #35782
• In _convert_delta_to_message_chunk, a try/except KeyError: pass wrapped the entire list comprehension for tool call chunks. If any single item in raw_tool_calls was missing the function or index key, the whole batch was silently discarded.
• Move the try/except inside the loop so only the malformed item is skipped — valid siblings are preserved.

Root cause

# before: one bad item drops everything
try:
    tool_call_chunks = [tool_call_chunk(...) for rtc in raw_tool_calls]
except KeyError:
    pass  # entire batch lost

# after: bad items are skipped individually
for rtc in raw_tool_calls:
    try:
        tool_call_chunks.append(tool_call_chunk(...))
    except KeyError:
        pass  # only this item skipped

Tests

Two new unit tests in test_base.py:

• test__convert_delta_to_message_chunk_malformed_tool_call — confirms the valid chunk survives when a sibling is missing function
• test__convert_delta_to_message_chunk_all_valid_tool_calls — confirms all chunks are returned when none are malformed

Areas requiring careful review

The except KeyError: pass behavior is intentional for genuinely incomplete streaming deltas (e.g. the first chunk may arrive with only index and no function yet). The per-item handling preserves that: a delta with only one incomplete item no longer poisons the rest.

This PR was developed with the assistance of an AI agent (Claude Sonnet 4.6).

Changed files

• libs/core/langchain_core/runnables/retry.py (modified, +8/-2)
• libs/core/tests/unit_tests/runnables/test_runnable.py (modified, +146/-0)
• libs/langchain_v1/langchain/agents/middleware/tool_selection.py (modified, +18/-5)
• libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_tool_selection.py (modified, +80/-0)
• libs/partners/openai/langchain_openai/chat_models/base.py (modified, +11/-11)
• libs/partners/openai/tests/unit_tests/chat_models/test_base.py (modified, +62/-0)
• libs/text-splitters/langchain_text_splitters/konlpy.py (modified, +7/-13)
• libs/text-splitters/langchain_text_splitters/nltk.py (modified, +4/-9)
• libs/text-splitters/langchain_text_splitters/sentence_transformers.py (modified, +7/-12)
• libs/text-splitters/langchain_text_splitters/spacy.py (modified, +12/-15)
• libs/text-splitters/tests/unit_tests/test_text_splitters.py (modified, +20/-0)

PR #417: feat(nodes): add llm_vision_ollama node for open-source image-to-text

• Repository: rocketride-org/rocketride-server
• Author: asclearuc
• State: closed | merged: True
• Link: https://github.com/rocketride-org/rocketride-server/pull/417

Description (problem / solution / changelog)

Summary

Implements a new vision node that uses locally-hosted open-source multimodal models via Ollama as an alternative to proprietary vision APIs. Supports LLaVA, Llama 3.2 Vision, Moondream, MiniCPM-V, and Qwen 2.5 VL models (10 pre-configured profiles). No API key required — runs fully on-premise.

Type

feat

Testing

• Tests added or updated
• Tested locally
• ./builder test passes

Checklist

• Commit messages follow conventional commits
• No secrets or credentials included
• Wiki updated (if applicable)
• Breaking changes documented (if applicable)

Linked Issue

Fixes #393

Summary by CodeRabbit

• New Features

  • Ollama Vision node for image-to-text analysis using local vision-capable models.
  • Supports chunked image uploads, automatic image encoding, and streaming image handling.
  • Many pre-configured model profiles plus a customizable profile.

• Improvements

  • UI exposes vision system/prompt editing, model selection, and local server settings.
  • More robust LLM calls with retries and clearer, user-friendly error messages.

Changed files

• nodes/src/nodes/llm_vision_ollama/IGlobal.py (added, +48/-0)
• nodes/src/nodes/llm_vision_ollama/IInstance.py (added, +81/-0)
• nodes/src/nodes/llm_vision_ollama/__init__.py (added, +39/-0)
• nodes/src/nodes/llm_vision_ollama/ollama_vision.py (added, +133/-0)
• nodes/src/nodes/llm_vision_ollama/requirements.txt (added, +9/-0)
• nodes/src/nodes/llm_vision_ollama/services.json (added, +390/-0)

PR #36662: fix(openai): handle malformed tool calls per-item instead of per-batch

• Repository: langchain-ai/langchain
• Author: billykailu1
• State: closed | merged: False
• Link: https://github.com/langchain-ai/langchain/pull/36662

Description (problem / solution / changelog)

Summary

Fixes #35782

• The _convert_delta_to_message_chunk function wrapped an entire list comprehension in a single try/except KeyError: pass block. If any tool call chunk in a streaming batch was malformed (e.g., missing "function" key), all tool call chunks — including valid ones — were silently dropped.
• This PR restructures the loop so each tool call chunk is processed independently: a KeyError in one item skips only that item, preserving all valid tool calls in the batch.
• Added a unit test (test_convert_delta_to_message_chunk_mixed_tool_calls) that verifies mixed valid/invalid tool call entries are handled correctly.

Areas requiring careful review

• The change is in libs/partners/openai/langchain_openai/chat_models/base.py around line 417. The logic is intentionally kept minimal — only the try/except scope was narrowed.

Test plan

• New unit test passes with mixed valid/invalid tool call chunks
• Full test suite passes (318 tests)
• Lint and mypy pass

This PR was developed with the assistance of an AI coding agent.

🤖 Generated with Claude Code

Changed files

• libs/partners/openai/langchain_openai/chat_models/base.py (modified, +12/-12)
• libs/partners/openai/tests/unit_tests/chat_models/test_base.py (modified, +30/-0)
• libs/partners/openai/uv.lock (modified, +3/-2)