firebase-ai-logic（Firebase AI logic 搜索）

Firebase AI Logic

When to use this skill

• Add Gemini-powered features directly inside a Firebase app.
• Build in-app text, chat, multimodal, structured-output, or streaming experiences.
• Decide whether a request should stay in the client/app layer or move to a backend workflow.
• Add Firebase-specific production controls such as App Check, quota awareness, and Firebase service integration around the AI feature.
• Handle app-side AI for web, mobile, Flutter, Unity, or similar Firebase-supported client surfaces.

Do not use this skill when

• The user needs server-owned flows, tool calling, RAG, evaluators, traces, prompt files, or reusable backend orchestration → use genkit.
• The user needs Firebase project setup, emulator, hosting/functions deploys, or CLI operations → use firebase-cli.
• The user is not committed to Firebase and mainly needs a generic provider SDK comparison → use a provider-specific or framework-selection skill instead.

Operating model

Treat firebase-ai-logic as the client/app integration anchor in a three-way lane:

1. App/client feature integration → firebase-ai-logic
2. Backend workflow orchestration → genkit
3. Firebase platform / operator tasks → firebase-cli

Do not blur these roles. If the request spans more than one role, split the packet and route each part explicitly.

Instructions

Step 1: Classify the request before giving implementation advice

Choose one mode first.

Mode A — Direct feature fit

Use this when the user wants to add an in-app AI feature such as:

• chat or assistant UI
• summarization or rewriting in the app
• multimodal prompts from user content
• structured output for app workflows
• streaming model responses into the UI

Deliver:

• whether Firebase AI Logic is the right lane
• what belongs in the app layer vs what should move out
• the minimum Firebase dependencies/services involved

Mode B — App wiring and UX integration

Use this when the user already chose Firebase AI Logic and needs help wiring it into the app.

Focus on:

• where the model call lives in the app architecture
• request/response lifecycle in UI state
• loading/error/retry/fallback UX
• prompt/template ownership in the app
• content moderation / safety surface the app must expose

Mode C — Production hardening

Use this when the feature exists or is close to launch.

Focus on:

• App Check / abuse prevention
• quota, cost, and rate-limit implications
• monitoring and failure visibility
• remote rollout / feature-flag strategy if relevant
• privacy boundaries for user input and generated output

Mode D — Escalation boundary

Use this when the user started with a client-side request, but the shape now implies backend orchestration.

Escalate to genkit when you see:

• tool calling or agent loops
• retrieval / RAG / database-grounded generation
• reusable flows shared across platforms
• evaluation harnesses, traces, or systematic observability
• secrets or privileged APIs that should not sit in the client

Step 2: Normalize the request packet

Rewrite the request into this brief before answering:

## Firebase AI Logic packet
- App surface: web | iOS | Android | Flutter | Unity | mixed
- User-facing feature: chat | summary | generation | multimodal | structured output | streaming
- Data involved: user text | images | docs | app state | Firebase data
- Safety/privacy concern: none | moderate | high
- Production stage: prototype | internal beta | launch prep | live issue
- Boundary check: stays in app | split with backend | move to `genkit`

If the packet says split with backend or move to genkit, say so immediately instead of pretending one skill owns the entire solution.

Step 3: Give app-layer guidance, not a giant code dump

For Firebase AI Logic requests, structure the answer in this order:

1. Why this should stay in the app layer
2. What to wire in the client UI/runtime
3. What Firebase-specific controls matter
4. What not to put in the client
5. When to escalate to genkit or firebase-cli

Use short code or pseudo-structure only when it clarifies the integration. Prefer architecture and boundary guidance over long SDK snippets.

Step 4: Cover the Firebase-specific concerns

Always touch the concerns that make this lane distinct from raw provider SDK usage:

• App Check / abuse prevention when relevant
• quotas / cost / rate awareness
• Firebase service interplay (Auth, Remote Config, analytics/monitoring, storage, functions handoff if needed)
• platform support implications for the target app surface
• rollout strategy for model changes or feature exposure when the request is production-facing

Step 5: Enforce honest route-outs

Use these route-outs explicitly.

Step 6: Output format

Respond with a compact packet like this:

## Recommendation
- Lane: Firebase AI Logic | Split with Genkit | Not a Firebase AI Logic fit
- Why: ...

## App-layer plan
1. ...
2. ...
3. ...

## Firebase-specific controls
- App Check / abuse:
- Quota / cost:
- Monitoring / rollout:

## Route-outs
- `genkit`: ...
- `firebase-cli`: ...

Best practices

1. Keep simple user-facing AI in the app layer only while it remains low-privilege and low-orchestration.
2. Move backend-worthy behavior out early; do not bolt tools/RAG/evals onto a client-only design.
3. Treat App Check, quota, and launch-readiness as first-class concerns, not afterthoughts.
4. Prefer one clear feature packet per response instead of mixing app UX, backend orchestration, and Firebase ops into a single blob.
5. When uncertain, explain the split between app-side Firebase AI Logic and backend-side genkit before proposing implementation steps.

Examples

Example 1: Should stay in Firebase AI Logic

Input: "Add streaming Gemini summaries inside our Firebase web app for article previews."

Output shape:

• lane = Firebase AI Logic
• plan focuses on app UI state, streaming rendering, prompt ownership, failure states, and App Check/quota guardrails
• route-out notes that reusable summarization pipelines or evaluation harnesses belong in genkit

Example 2: Should split with Genkit

Input: "Our mobile app needs chat, tool calling, and retrieval over Firebase data with server-side observability."

Output shape:

• lane = Split with Genkit
• Firebase AI Logic stays at the app interaction edge only if needed
• backend workflow, retrieval, tools, and observability move to genkit

References

• Firebase AI Logic overview
• Firebase AI Logic get started
• Firebase AI Logic production checklist
• Genkit docs