Elixir LangChain v0.6.3 is out! This release focuses on stability across providers and new capabilities that make building with #LLMs in Elixir more reliable. github.com/brainlid/lan... #ElixirLang #MyElixirStatus Here's what's new and why it matters. ๐Ÿงต

GitHub - brainlid/langchain: E...
GitHub - brainlid/langchain: Elixir implementation of a LangChain style framework that lets Elixir projects integrate with and leverage LLMs.

Elixir implementation of a LangChain style framework that lets Elixir projects integrate with and leverage LLMs. - brainlid/langchain

GitHub
Streaming with reasoning models (Qwen3, DeepSeek-R1) could crash with a CaseClauseError. Their `<think>` tokens produced empty responses we didn't handle. Now returns a clean error instead of crashing.
Two provider streaming crashes fixed: - Azure OpenAI keepalive SSE events no longer crash the parser - Mistral responses without tool_calls no longer hit the error handler
Anthropic requires `additionalProperties: false` on tool schemas or you get a cryptic `invalid_request_error`. LangChain now adds this automatically.
ChatOpenAI now supports `logprobs` and `top_logprobs`. Probability data shows up in message metadata for both streaming and non-streaming. Great for evals and confidence scoring.
ChatVertexAI tool results now support images, PDFs, and other base64 files - not just text. Multimodal agent workflows on Google's platform.
New: `LangChain.Trajectory` for agent evaluation. Structured inspection of agent execution paths - what it did, in what order, and whether it made the right calls. Upgrade: {:langchain, "~> 0.6.3"} hexdocs.pm/langchain/La...

LangChain.Trajectory โ€” LangCha...
LangChain.Trajectory โ€” LangChain v0.6.3