Build production-grade Firebase Genkit applications including RAG systems, multi-step flows, and tool-calling agents for Node.js, Python, and Go. This skill covers the full lifecycle from project scaffolding and schema validation through flow implementation, local testing with the Genkit Developer UI, and deployment to Firebase Functions or Cloud Run with AI monitoring and OpenTelemetry tracing.
npm install genkit @genkit-ai/googleai for TypeScript)npm install -g firebase-tools)tsconfig.json, genkit.config.ts, or equivalent)ai.defineFlow() with model configuration, temperature tuning, and token limitsai.defineTool() with scoped schemas for each external capability the flow requiresai.defineRetriever() with embedding generation (text-embedding-gecko) and vector database integrationSAFETY_BLOCK), quota exceeded (QUOTA_EXCEEDED), and provider timeoutsSee ${CLAUDE_SKILL_DIR}/references/how-it-works.md for the phased workflow and ${CLAUDE_SKILL_DIR}/references/production-best-practices-applied.md for the production checklist.
firebase.json) or Cloud Run service YAML| Error | Cause | Solution |
|---|---|---|
SAFETY_BLOCK response |
Model safety filters triggered on input or output | Review prompt content; adjust safety settings; add input sanitization before generation |
QUOTA_EXCEEDED |
API rate limit or daily token quota reached | Implement exponential backoff with jitter; request quota increase; cache repeated prompts |
| Schema validation failure | Runtime input does not match Zod/Pydantic schema | Add descriptive error messages to schema; validate inputs before calling ai.generate() |
| Retriever returns empty results | Vector database query found no matches above similarity threshold | Lower similarity threshold; verify embeddings are indexed; check embedding model version match |
| Deployment timeout | Cold start exceeds Firebase Functions 60s limit | Increase memory allocation; use Cloud Run for long-running flows; enable min instances > 0 |
See ${CLAUDE_SKILL_DIR}/references/errors.md for additional error scenarios.
Scenario 1: Question-Answering Flow -- Create a Genkit flow using Gemini 2.5 Flash with Zod input/output schemas. Set temperature to 0.3 for factual responses. Deploy to Firebase Functions with token usage monitoring. Expected latency: under 2 seconds per query.
Scenario 2: RAG Document Search -- Implement a retriever with text-embedding-gecko embeddings connected to Firestore vector search. Build a RAG flow that retrieves top-5 relevant documents, injects them as context, and generates grounded answers with source citations. Include context caching for repeated queries.
Scenario 3: Multi-Tool Agent -- Define weather and calendar tools with typed schemas. Create an agent flow that routes user queries to appropriate tools, handles multi-turn conversations, and traces each tool execution for debugging. Deploy to Cloud Run with auto-scaling (2-10 instances).
See ${CLAUDE_SKILL_DIR}/references/workflow-examples.md for complete code examples.