Database-Driven Intelligence Providers — Design

Goal

Replace env-var-based provider configuration with a database-driven IntelligenceProvider model (same pattern as NotificationChannel). Add 6 new AI provider drivers (Claude, Gemini, Copilot, Grok, Ollama, Mistral) alongside the existing OpenAI driver. The local provider remains the fallback/default when no AI provider is configured.

Architecture

Pipeline triggers intelligence stage
        ↓
DB has active AI provider?
    YES → AI provider (from DB) → incident → LLM API → Recommendations
    NO  → Local provider (fallback) → rule-based analysis → Recommendations
        ↓
Recommendations stored in AnalysisRun

Key Decisions

1. DB-driven config — IntelligenceProvider model stores provider type, credentials (JSONField), active flag. No env vars required (but OpenAI backward-compat preserved).
2. Single active provider — One provider is active at a time. is_active flag with DB constraint.
3. Local as fallback — Local provider runs when no AI provider is active in the DB. It provides rule-based recommendations without requiring any external API.
4. Official SDKs — Each provider uses its official Python SDK with lazy imports (conditional on package availability).
5. Same prompt strategy — All AI providers use the same system prompt and incident formatting as OpenAI. Only the API call differs.

Model: IntelligenceProvider

class IntelligenceProvider(models.Model):
    name = models.CharField(max_length=100, unique=True)
    provider = models.CharField(max_length=50, db_index=True)  # claude, gemini, etc.
    config = models.JSONField(default=dict, blank=True)  # {api_key, model, max_tokens, ...}
    is_active = models.BooleanField(default=False, db_index=True)
    description = models.TextField(blank=True, default="")
    created_at = models.DateTimeField(auto_now_add=True)
    updated_at = models.DateTimeField(auto_now=True)

Config JSONField per provider:

Provider	Required Config Keys	Optional Keys
openai	api_key	model (gpt-4o-mini), max_tokens (1024)
claude	api_key	model (claude-sonnet-4-20250514), max_tokens (1024)
gemini	api_key	model (gemini-2.0-flash), max_tokens (1024)
copilot	api_key	model (gpt-4o), max_tokens (1024), base_url
grok	api_key	model (grok-3-mini), max_tokens (1024), base_url (https://api.x.ai/v1)
ollama	—	host (http://localhost:11434), model (llama3.2), max_tokens (1024)
mistral	api_key	model (mistral-small-latest), max_tokens (1024)

Provider Drivers

Base Changes

Refactor existing BaseProvider to support:

• Receiving config from DB (via provider_config dict)

New Abstract: BaseAIProvider(BaseProvider)

Intermediate class for all LLM-backed providers:

• _build_prompt(incident) — shared prompt construction
• _parse_response(response) — shared JSON response parsing (already in OpenAI, extract to base)
• _get_fallback_recommendation(incident, error_message) — shared fallback

Provider Implementations

File	Class	SDK	Notes
openai.py	OpenAIRecommendationProvider	openai	Refactor to use BaseAIProvider
claude.py	ClaudeRecommendationProvider	anthropic	Messages API
gemini.py	GeminiRecommendationProvider	google-genai	GenerativeModel API
copilot.py	CopilotRecommendationProvider	openai	OpenAI-compatible endpoint
grok.py	GrokRecommendationProvider	openai	OpenAI-compatible endpoint (x.ai)
ollama.py	OllamaRecommendationProvider	ollama	Local model, no API key needed
mistral.py	MistralRecommendationProvider	mistralai	Mistral chat API

Registration & Selection

# providers/__init__.py
PROVIDER_CLASSES: dict[str, type[BaseProvider]] = {
    "local": LocalRecommendationProvider,
    "openai": OpenAIRecommendationProvider,
    # ... conditionally registered based on SDK availability
}

def get_active_provider(**kwargs) -> BaseProvider:
    """Get the active AI provider from DB, fall back to local."""
    from apps.intelligence.models import IntelligenceProvider as ProviderModel
    try:
        db_provider = ProviderModel.objects.filter(is_active=True).first()
        if db_provider and db_provider.provider in PROVIDER_CLASSES:
            cls = PROVIDER_CLASSES[db_provider.provider]
            return cls(**db_provider.config)
    except Exception:
        pass
    return LocalRecommendationProvider()

Admin

Extend intelligence admin with:

• IntelligenceProviderAdmin — list display, config (redacted in display), active toggle
• Validate only one provider is active (or none = local-only mode)

Migration Path

• Existing OPENAI_* env vars still work as defaults when no DB record exists
• get_provider() checks DB first, falls back to env-var-based OpenAI, then local
• No breaking changes to orchestration layer — it still calls get_provider() / provider.run()