feat(chat): add multi-provider chat/LLM abstraction (#285)

[Kaap10]

Description

Implements a unified multi-provider chat/LLM abstraction layer mirroring Dynavec's embedder architecture. This provides a single consistent interface to interact with multiple LLM backends (OpenAI, Anthropic, and AWS Bedrock) with full support for message formatting, function/tool calling, structured output schemas, and token streaming.

Related issue

Fixes #285

Changes

• Core Chat Abstraction (src/dynavec/chat/base.py):
  • Defined standard data models: Message, Role, Tool, ToolCall, ChatResult, and ChatChunk.
  • Defined abstract base class ChatModel providing sync (invoke, stream) and async (ainvoke, astream) execution interfaces.
• Provider Implementations:
  • src/dynavec/chat/openai.py: Added OpenAIChatModel with OpenAI function/tool calling and streaming delta support.
  • src/dynavec/chat/anthropic.py: Added AnthropicChatModel mapping conversation roles to Anthropic content blocks, handling system prompts, tool uses, and event streaming.
  • src/dynavec/chat/bedrock.py: Added BedrockChatModel powered by the unified AWS Bedrock Converse and ConverseStream APIs.
• Module Exports (src/dynavec/chat/__init__.py):
  • Configured lazy module imports for provider classes to keep external SDK dependencies (openai, anthropic, boto3) optional.
• Testing (tests/test_chat_models.py):
  • Added 6 comprehensive offline unit tests covering invoke, tool-calling, stream handling, async delegation, and missing dependency guards across all providers.

Testing

• Tests pass locally
• Ruff checks pass
• Documentation updated, if applicable

Checklist

• My changes are focused and relevant to this pull request.
• I have added or updated tests where appropriate.
• I have reviewed my changes for unrelated modifications.
• I have updated documentation where necessary.

[Kaap10]

Ready for review!

[codeforstartups]

Reviewed in depth — this is a strong foundation for the dynaflow model layer (closes #285). 🎯

Design — the ChatModel ABC cleanly mirrors our embedder abstraction: sync invoke/stream, async ainvoke/astream defaulting to asyncio.to_thread, and clean dataclasses (Message / Role / Tool / ToolCall / ChatResult / ChatChunk). Function/tool calling and structured output are modeled from the start.

Optional deps done right — provider SDKs (openai / anthropic / boto3) stay optional via lazy __getattr__ imports in chat/__init__.py; anthropic added as its own extra + to all/typecheck + mypy overrides.

Tests — 6 genuine offline tests with per-provider fakes, async delegation, exports, and missing-dependency guards. Verified locally (6 passed) and all 4 CI checks pass incl. strict mypy on 3.9/3.11/3.12.

Two non-blocking follow-ups for later (not needed to merge):

1. The default astream buffers the full sync stream into a list before yielding, so the async fallback loses incremental streaming — providers can override for true async streaming (Anthropic/OpenAI both support it).
2. Gemini/Vertex + Azure OpenAI remain to round out the provider matrix from Multi-provider chat/LLM abstraction (Bedrock, OpenAI, Anthropic, Gemini/Vertex, Azure OpenAI, local) #285.

Merging — excellent work @Kaap10!