Pluggable Multi-Provider LLM Diagnostics

Triage features a pluggable, multi-provider LLM architecture. Whether your team relies on cloud frontier models (Google Gemini, OpenAI, Anthropic Claude) or runs 100% air-gapped on-premise local models (Ollama, vLLM, LocalAI, LM Studio), Triage delivers structured incident diagnostics, unified git diff generation, and clean source-level fix application.

---

Supported AI Providers

Provider	Supported Models	Connection Type
Google Gemini	gemini-2.0-flash, gemini-2.5-pro, gemini-1.5-flash	Official Google GenAI SDK
OpenAI	gpt-4o, gpt-4o-mini, o3-mini, o1, gpt-4.5-preview	OpenAI Chat Completions API
Anthropic Claude	claude-3-5-sonnet-20241022, claude-3-7-sonnet-20250219, claude-3-5-haiku	Claude Messages API
Local / Ollama	deepseek-coder-v2, qwen2.5-coder:7b, llama3.3	OpenAI-compatible Base URL (http://localhost:11434/v1)

---

Air-Gapped & Local Model Setup (Ollama / vLLM)

For high-security or on-premise deployments requiring zero outbound telemetry:

1. Start your local Ollama instance with a code-specialized model:

   ollama run deepseek-coder-v2

2. In the Triage Studio Dashboard, navigate to Settings > AI Configuration (or Step 4 of the Setup Wizard).
3. Select Local / Ollama:
   • Base URL: http://localhost:11434/v1 (or your vLLM / LocalAI endpoint)
   • Model Name: deepseek-coder-v2 or qwen2.5-coder:7b
   • API Key: (leave blank if authentication is not required)
4. Click Test Connection to benchmark endpoint latency and verify model availability before saving.

---

Structured Output Schema

The Triage Engine guarantees strict JSON schema output across all providers:

{
  "type": "object",
  "properties": {
    "root_cause": {
      "type": "string",
      "description": "Precise explanation of why the Go runtime panicked on the triggering line."
    },
    "suggested_fix": {
      "type": "string",
      "description": "Exact code correction required to prevent the panic."
    },
    "severity": {
      "type": "string",
      "enum": ["CRITICAL", "HIGH", "MEDIUM", "LOW"]
    },
    "suggested_patch": {
      "type": "string",
      "description": "Unified Git diff format patch."
    }
  },
  "required": ["root_cause", "suggested_fix", "severity"]
}

---

Domain-Aware AI Triage & Project Context

To ensure LLMs diagnose crashes accurately without violating business rules, Triage supports optional Project & Architectural Domain Context.

How It Works

1. During project onboarding (or later in Settings > General), specify domain constraints such as ledger idempotency rules, state machine transitions, concurrency invariants, or database transaction semantics.
2. When a panic occurs, the engine automatically injects your domain context into the prompt:

   ### Project & Domain Context:
   
   High-throughput payment gateway processing Stripe and crypto webhooks with strict ledger idempotency and database transaction rollbacks.

3. The active provider uses this context in AnalyzeCrash, GeneratePatch, and ApplyFixToFile to synthesize patches that strictly adhere to your business boundaries.

---

Automated File Patching (ApplyFixToFile)

When generating a bugfix Pull Request, the engine uses the configured LLM’s code reasoning to merge the suggested patch into the latest source file retrieved from GitHub:

1. Preserves existing package headers, imports, comments, and project indentation.
2. Applies necessary nil guards, error validations, or constructor initializations.
3. Automatically formats the resulting Go code with standard go/format before committing directly to the bugfix branch (triage/fix-...).

---

Direct REST Endpoints

The Triage engine provides direct HTTP endpoints for on-demand AI operations:

1. Test Provider Connection

curl -X POST http://localhost:8080/api/v1/settings/llm/test \
  -H "Content-Type: application/json" \
  -d '{
    "provider": "openai",
    "api_key": "sk-...",
    "model": "gpt-4o"
  }'

2. Panic Root Cause Analysis

curl -X POST http://localhost:8080/api/v1/llm/analyze-panic \
  -H "Content-Type: application/json" \
  -d '{
    "triggeringFile": "handlers/payment.go",
    "panicMessage": "runtime error: invalid memory address or nil pointer dereference",
    "astCode": "func ProcessTransaction(w http.ResponseWriter, r *http.Request) {\n..."
  }'

3. Git Diff Patch Generation

curl -X POST http://localhost:8080/api/v1/llm/generate-patch \
  -H "Content-Type: application/json" \
  -d '{
    "triggeringFile": "handlers/payment.go",
    "panicMessage": "runtime error: invalid memory address or nil pointer dereference",
    "astCode": "func ProcessTransaction(w http.ResponseWriter, r *http.Request) {\n...",
    "rootCause": "Attempted to evaluate req.Amount on an uninitialized nil pointer on line 28."
  }'

Dynamic Configuration & Key Management

AI provider credentials, models, and custom endpoint URLs are managed dynamically through the Setup Wizard and Studio Dashboard (Settings > AI Configuration).

• Database-Backed: Settings are stored securely in the embedded database (instance_config table).
• Hot-Swappable: You can switch providers (e.g. from Gemini to local Ollama or Claude) or update models on the fly without restarting the engine container.
• Connection Testing: The built-in /api/v1/settings/llm/test endpoint lets you verify authentication and benchmark latency before saving changes.