Skip to content

CHANGELOG.md

Latest commit

 

History

History
78 lines (57 loc) · 3.38 KB

CHANGELOG.md

File metadata and controls

78 lines (57 loc) · 3.38 KB
Raw

Changelog

All notable changes to ollama_client_rs are documented here.

[0.3.0] — 2025-04-15

Added

Model-aware tool calling — the crate now understands the different tool-call formats used by each model family and selects the correct one automatically.

Type Details
ModelFamily Enum covering Gemma, Qwen, Llama, Mistral, DeepSeek, Phi, CommandR, Unknown
ToolFormat Enum: Native, PromptInjectedJson, HermesXml, NativeFunctionTag
ModelInfo Struct with family, tool_format, context_length, supports_thinking
ModelInfo::from_name() Heuristic detection from model name string
OllamaClient::model_info() Enriches heuristics with live Ollama API data

ChatRequestBuilder — fluent builder API for constructing chat requests:

  • Automatically injects tools into the system prompt for Gemma (prompt-injected format)
  • Passes tools as native Ollama tools field for all other families
  • Helper constructors: ChatRequest::new(), ChatRequest::builder()

Message helper constructors — Message::system(), Message::user(), Message::assistant(), Message::tool_result() for ergonomic message building.

Tool::function() — convenience constructor for function-type tools.

ChatResponse::extract_tool_calls() — smart extraction that:

  1. Returns native tool_calls field if populated
  2. Falls back to parsing content for Gemma JSON format
  3. Falls back to parsing Hermes <tool_call> XML tags
  4. Falls back to parsing Llama <function=name> tags

OllamaClient::pull_model() — stream model downloads with progress events.

Streaming improvements:

  • Buffer now handles chunk boundaries that fall mid-JSON-line (previously could panic)
  • Synthetic done=true chunk emitted if server closes connection without one
  • Tool calls are normalised on the final streaming chunk automatically

Automatic tool-call normalisation — both chat() and chat_stream() now call extract_tool_calls() on the response and populate message.tool_calls if it was empty but content contained parseable tool calls. Callers no longer need to check both fields.

New examples

Example Description
basic List models, simple chat, structured JSON output, model info
streaming Token-by-token streaming with usage statistics
tool_calling Full two-turn tool-use loop for native models (Qwen, Mistral)
gemma_tools Prompt-injected tool calling for Gemma with content parsing
multi_turn Interactive REPL with full conversation history
embeddings Batch embeddings, cosine similarity, semantic search
model_info Model family detection and capability overview
pull_model Download a model with a progress bar
agentic_loop Full ReAct-style agent loop with tool dispatch

Changed

  • ChatRequest now derives Clone to support multi-turn conversation patterns
  • EmbedInput supports both Single(String) and Multiple(Vec<String>) variants
  • Cargo.toml: added keywords and categories for crates.io discoverability

Fixed

  • Streaming parser no longer fails when a TCP chunk boundary falls mid-JSON-line
  • chat() and chat_stream() now correctly report tool calls for models that embed them in content rather than the native tool_calls field

[0.2.0]

Initial public release with basic chat, streaming, and embeddings support.