langchain-workflow-provider-agnostic

Langchain Workflow — Provider-Agnostic Implementation

Changes to the LangGraph workflow and LLM service must work regardless of which model/provider is chosen (lmstudio, openai, anthropic).

Rules

1. No provider-specific branching — Do not add logic that only runs for one provider (e.g. if provider == "openai" for behavior changes). Use the existing LLMService abstraction and _canonicalize_provider for routing.

2. Use config-driven provider/model — Read provider and model from config_models or config_snapshot; never hardcode a provider or model name in workflow logic.

3. Prompt and schema parity — Prompts, output schemas, and parsing must work across all providers. Avoid provider-specific prompt variants or response formats.

4. Parameter bounds — Temperature, max_tokens, and other inference params must use provider-agnostic defaults or provider-specific defaults from provider_defaults; do not assume LMStudio-only semantics.

Examples

# ❌ BAD — provider-specific behavior
if provider == "lmstudio":
    messages = build_messages_v1(prompt)
else:
    messages = build_messages_v2(prompt)

# ✅ GOOD — single path for all providers
messages = build_messages(prompt)

# ❌ BAD — hardcoded model
model = "gpt-4o-mini"

# ✅ GOOD — from config/env
model = config_models.get("RankAgent") or self.provider_defaults.get(provider)