[bot] Add instrumentation support for kalosm local/cloud AI toolkit

[braintrust-bot]

Summary

The Braintrust Rust SDK has no instrumentation support for kalosm, a Rust-native AI toolkit backed by HuggingFace that provides high-level ChatModel, TextCompletionModel, and Embedder execution traits for both local (Llama, Bert) and cloud-compatible (OpenAI-compatible, Anthropic-compatible) backends. No other Braintrust SDK instruments this library, as it is Rust-specific, but Braintrust documents HuggingFace as a supported TypeScript integration.

What is missing

kalosm::language defines clear execution-oriented traits that could be instrumented to automatically capture spans:

• ChatModel / ChatModelExt — chat() async method for multi-turn conversation, with streaming via CreateChatSession
• TextCompletionModel / TextCompletionModelExt — stream_text() for streaming text generation
• StructuredChatModel / StructuredTextCompletionModel — structured/typed output generation
• Embedder / EmbedderExt — embed() for converting text to embedding vectors

A wrapper or span helper around these traits would enable automatic tracing of kalosm-based Rust applications, capturing inputs, outputs, latency, time-to-first-token, and (where available) token counts for any model backend kalosm supports.

The SDK already has building blocks that could support this:

• wrap_stream_with_span() in src/stream.rs for streaming instrumentation
• SpanBuilder / SpanHandle in src/span.rs for manual span creation

But there is no integration that connects these to kalosm's execution traits.

Library significance

• 2.2k GitHub stars in the floneum monorepo (44 open issues, active development)
• v0.4.0 stable release
• Supports local model backends: Llama (text generation), Bert (embeddings), Whisper (audio), Stable Diffusion (image)
• Supports cloud-compatible endpoints: OpenAI-compatible and Anthropic-compatible APIs via the same interface
• Includes RAG workflows via vector stores (SurrealDB integration), document chunking, and semantic search
• Official docs: https://docs.rs/kalosm/latest/kalosm/

Braintrust docs status

not_found — kalosm is not listed in Braintrust's SDK integrations, AI providers page, or trace LLM calls documentation. Braintrust does list Hugging Face as a supported TypeScript integration (https://www.braintrust.dev/docs/instrument/trace-llm-calls), and kalosm is the primary HuggingFace-backed local inference framework for Rust.

• https://www.braintrust.dev/docs/instrument/trace-llm-calls — lists Hugging Face for TypeScript but no Rust equivalent
• https://www.braintrust.dev/docs/integrations/ai-providers — no mention of kalosm or local Rust inference

Upstream sources

• GitHub repository: https://github.com/floneum/floneum/tree/main/interfaces/kalosm
• API documentation: https://docs.rs/kalosm/latest/kalosm/
• kalosm::language API (ChatModel, Embedder traits): https://docs.rs/kalosm/latest/kalosm/language/index.html
• Crate page: https://crates.io/crates/kalosm

Local files inspected

• src/extractors.rs — only extract_openai_usage() and extract_anthropic_usage() exist; no kalosm-related code
• src/stream.rs — wrap_stream_with_span() is generic over Stream<Item = Result<Value, E>>; kalosm streams yield tokens directly rather than JSON chunks, requiring a different adapter
• src/lib.rs — public API exports; no kalosm references
• Cargo.toml — no kalosm dependency
• Full codebase grep for kalosm, ChatModel, TextCompletionModel, Embedder — zero results