TL;DR

LangChain has been updated with improved streaming support for content blocks and enhanced HumanInTheLoopMiddleware for better control over agent interactions. Several dependency updates and fixes across integrations (Anthropic, Deepseek, Perplexity) improve stability and functionality.

Breaking

None

New

• Content Block Streaming: Added version 2 of content-block-centric streaming for more efficient data handling (core).
• HITL Control: New interrupt_mode and when predicate added to HumanInTheLoopMiddleware (human-in-the-loop).
• Model Profiles: Expanded model profiles now include text_inputs and text_outputs for richer metadata.

Fixes Worth Knowing

• Resolved issues with URL parsing for Deepseek Azure endpoints.
• Fixed tool choice logic for Azure deployments in Deepseek.
• Normalized tool-call IDs across providers in Anthropic integration.
• Improved guidance for rejecting human-in-the-loop suggestions.
• Fixed missing fields and schema drift warnings in Model Profiles.

Before You Upgrade

Ensure your langchain-core dependency is at least version 1.3.3 for optimal compatibility.

TL;DR

Langchain now supports tool calling with Anthropic models, enabling more complex agent interactions and expanding the range of supported LLMs.

Breaking

• LLMChain, RetrievalQA, and ConversationalRetrievalChain are deprecated.
• Some chains have removed hidden defaults, requiring more explicit configuration.
• The tool.__call__ method is deprecated.

New

• Upstage Integration: Added support for the Upstage LLM provider.
• Graph Visualization Improvements: Enhanced the runnable graph visualization for better debugging.
• Hybrid Search: Added hybrid search functionality to CassandraVectorStore.

Fixes Worth Knowing

• Fixed issues with Ollama integration, including LLaMA3 support.
• Resolved problems with embedding and message formatting for Anthropic models on Amazon Bedrock.
• Corrected issues with token handling in Groq integration.
• Fixed a bug preventing agent_token_buffer_memory from working with OpenAI tools.

Before You Upgrade

• Review your code for usages of the deprecated chains (LLMChain, RetrievalQA, ConversationalRetrievalChain) and update to newer alternatives.
• If you are using tool

TL;DR

LangChain now supports tool calling across multiple providers (Anthropic, Groq, Mistral, OpenAI) and includes a new aformat feature for more flexible prompt templating.

Breaking

• The langchain-postgres package has been removed.

New

• Tool Calling: Added support for tool calling across multiple LLM providers, enabling agents to utilize tools more effectively.
• aformat: Introduced aformat for prompt templates, allowing for more dynamic and flexible prompt creation.
• New Integrations: Added integrations for Pebble, MLX, and updated integrations for Cohere and Together AI.

Fixes Worth Knowing

• Fixed issues with OpenAI chat model parameters and Azure embedding length checks.
• Resolved several bugs related to Anthropic integration, including message type lookups and structured output.
• Improved stability and fixed bugs in various community integrations (OpenSearch, Elasticsearch, Tencent Cloud VectorDB).
• Corrected issues with token counting in Bedrock and fixed a bug in Pinecone upsert functionality.

Before You Upgrade

• If you were using langchain-postgres, remove the dependency and migrate to an