rzyfront/vendix-ai-streaming

Source of Truth

• Backend stream controller: apps/backend/src/ai-engine/ai-stream.controller.ts
• Core stream method: apps/backend/src/ai-engine/ai-engine.service.ts
• Providers: apps/backend/src/ai-engine/providers/
• Angular stream service: apps/frontend/src/app/core/services/ai-stream.service.ts
• Text component: apps/frontend/src/app/shared/components/ai-text-stream/
• Chat SSE endpoint: apps/backend/src/domains/store/ai-chat/ai-chat.controller.ts

Backend Streaming

AIEngineService.runStream(appKey, variables?, extraMessages?) validates app/provider, runs subscription gate, checks rate limit, builds messages, requires provider chatStream(), yields chunks, consumes quota on final done, and logs in finally.

Chunk shape:

type AIStreamChunk =
  | { type: 'text'; content: string }
  | { type: 'done'; usage?: { promptTokens: number; completionTokens: number; totalTokens: number } }
  | { type: 'error'; error: string };

OpenAI-compatible provider streams text deltas and final done; usage may remain zero unless stream usage is explicitly requested. Anthropic provider uses client.messages.stream() and finalMessage() for usage.

SSE Endpoints

Generic AI endpoint:

• GET /store/ai/stream/:appKey.
• Query params become variables; token is stripped.
• Emits custom event type ai-chunk with JSON chunks.

Chat endpoint:

• SSE /store/ai-chat/conversations/:id/stream.
• Reads message content from query param content.
• Emits ai-chunk events and completes on done/error.

Frontend Streaming

AIStreamService.streamRun(appKey, variables?, token?) uses EventSource, listens to ai-chunk, parses JSON, closes on done, error, parser failure, or unsubscribe.

app-ai-text-stream accepts stream$: Observable<string> | null, appends emitted strings to displayText, and shows a cursor while streaming.

Current caveat: app-ai-text-stream cleans up on component destroy, but replacing stream$ is not a full old-stream cleanup pattern. Verify before using it for frequently swapped streams.

Chat Streaming Caveat

Chat reducer supports streaming chunk accumulation, but current AIChatEffects use sync HTTP send and do not dispatch streaming actions. Do not document chat SSE as active frontend behavior unless wiring is added.

Rules

• Always close EventSource on completion/error/unsubscribe.
• Complete Nest subscribers on done or error chunks.
• Keep logging in finally for backend streams.
• Do not use WebSocket for simple one-way AI token streams unless requirements change.

Related Skills

• vendix-ai-platform-core
• vendix-ai-chat
• vendix-notifications-system