Rationale

• Enable multi-cloud AI provider flexibility without touching agent code.
• Provide a unified interface to swap AI backends (OpenAI, AWS Bedrock, GCP Model Garden/Vertex).
• Reduce duplication and improve maintainability by centralizing provider selection and configuration.

Related issue: closes #3

Summary of Changes

• Core
  • Added src/core/ai.py with get_chat_model() factory supporting:
    • OpenAI (default) via langchain_openai
    • AWS Bedrock via langchain_aws (preferred) or langchain_community fallback
    • GCP Model Garden/Vertex via langchain_google_vertexai
  • Extended AIConfig in src/core/config.py with provider-specific fields:
    • Bedrock: bedrock_region, bedrock_model_id
    • GCP: gcp_project, gcp_location
  • Added validation for provider-specific requirements.
• Agents
  • Replaced direct ChatOpenAI usage with get_chat_model() in:
    • src/agents/base.py
    • src/agents/feasibility_agent/nodes.py
    • src/agents/engine_agent/nodes.py
    • src/agents/acknowledgment_agent/agent.py
• No behavioral changes to agent APIs; only the AI backend instantiation is abstracted.

Behavior Change

• Agents now obtain the chat model via a provider-agnostic factory.
• Switching providers requires only environment variable changes; no code edits.

Configuration

• Env vars:
  • Common:
    • AI_PROVIDER = openai | bedrock | garden
    • AI_MODEL, AI_MAX_TOKENS, AI_TEMPERATURE
  • OpenAI:
    • OPENAI_API_KEY
  • AWS Bedrock:
    • BEDROCK_REGION (optional hint)
    • BEDROCK_MODEL_ID (optional; falls back to AI_MODEL)
    • Credentials resolved via AWS env/IMDS
  • GCP Garden/Vertex:
    • GCP_PROJECT_ID (optional, recommended)
    • GCP_LOCATION (optional, recommended)
    • Authentication via ADC or env credentials

Optional dependencies (to be installed by users selecting these providers):

• Bedrock: langchain-aws (preferred) or langchain-community
• GCP/Vertex: langchain-google-vertexai

Performance

• No change to algorithmic complexity.
• Adheres to performance policy: static → hybrid → LLM. Only instantiation path changed.

Security

• No secrets are logged.
• Credentials are read from environment/standard cloud auth mechanisms.
• No changes to webhook verification or secret handling.

Risk and Rollback

• Risk: Import errors if optional provider deps are missing and AI_PROVIDER is set accordingly.
  • Mitigation: Clear runtime error messages guide installation.
• Rollback plan: Set AI_PROVIDER=openai to use the default path; code remains compatible.

CI

• Types and lint pass locally.
• No changes to CI workflows required.

PR Checklist (per CONTRIBUTING)

• Types added/updated; Pydantic at boundaries unchanged
• Tests green locally (lint/type/tests)
• Performance order respected (static→hybrid→LLM)
• Small, focused edits; no unrelated refactors
• No secrets in logs or prompts
• Docs updated (can be added to this PR if maintainers prefer)

Branch

• feat/ai-wrapper-bedrock-garden