Aller au contenu principal

Architecture: AI Facade

Status: Active Last updated: 2026-05-01 Audience: AI agents and engineers debugging AI provider routing, model selection, structured output, or streaming chat.

1. Purpose​

AiFacadeService is the single platform-side entry point for every LLM call the platform makes — generation pipelines, AI conversations, content extraction validation, comparison generation, item enrichment, and dashboard test calls. It resolves which AI provider plugin should handle the call, picks the right model tier for the task, applies provider-specific routing (reasoning, JSON mode, embedding shape), and provides graceful degradation across plugins.

Plugins themselves are OpenAI-shape-compatible — every AI provider plugin (openai, anthropic, google, groq, ollama, mistral, openrouter, vercel-ai-gateway, plus the various aggregator implementations) wraps a LangChain client behind the same IAiProviderPlugin contract. The facade handles which plugin and which model; the plugin handles the wire protocol.

2. Public Surface (IAiFacade)​

interface IAiFacade {
	chatCompletion(opts: ChatCompletionOptions): Promise<ChatCompletionResponse>;
	streamChatCompletion(opts: ChatCompletionOptions): AsyncIterable<ChatCompletionChunk>;
	askJson<T>(opts: AskJsonOptions<T>): Promise<AskJsonResponse<T>>;
	askJsonCompletion<T>(opts: AskJsonOptions<T>): Promise<AskJsonCompletionResponse<T>>;
	embeddings(opts: EmbeddingsOptions): Promise<EmbeddingsResponse>;
	listModels(opts: { providerId?: string }): Promise<AiModel[]>;
	getProviderConfig(providerId: string): Promise<AiProviderConfig>;
	validateConnection(providerId: string): Promise<ConnectionValidationResult>;
}

Every method accepts a FacadeOptions block carrying userId, workId?, optional explicit providerId, and routing overrides. Without these the facade falls back to the cascade (§3).

3. Provider Resolution Cascade​

The facade resolves a request to a plugin in this order:

  1. Explicit providerId in the options (admin tool calls, "test this provider" button).
  2. Per-work binding — work_plugins.aiProvider row.
  3. Per-user binding — user_plugins.aiProvider row.
  4. defaultFor: 'ai-provider' plugin — today: openrouter.
  5. Throw AiFacadeError("no provider") — caller handles.

The cascade is the same one defined in Plugin SDK §11 and shared with every other facade.

4. Model Tier Routing​

Every AI provider plugin declares four model aliases in its settings schema:

AliasTypical useDefault cost target
simpleModelTags, slugs, short labels, classificationsCheapest tier
mediumModelItem summaries, descriptions, reformattingMid tier
complexModelFull-page generation, multi-step analysis, reasoningTop tier
defaultModelUsed when no complexity is specifiedMid tier (typical)

The facade accepts a complexity: TaskComplexity hint ('simple' | 'medium' | 'complex') and resolves the right alias from the plugin's resolved settings. The plugin substitutes the alias for the actual model id (e.g. simpleModel: 'gpt-5-nano') before calling LangChain.

Pipelines pick complexity per step:

  • Domain detection → simple
  • Item summary → medium
  • Whole-page generation → complex
  • Reasoning (deep analysis) → complex with routing.reasoningEffort

5. Model Catalog​

fetchModelCatalog() and matchModelCatalogEntry() (in packages/agent/src/facades/model-catalog.ts) provide a provider-agnostic model registry the facade caches per process for one hour. The catalog tracks:

  • id — provider-namespaced model id (e.g. anthropic/claude-sonnet-4-5).
  • provider — the upstream provider.
  • family — model family (Claude Sonnet 4, GPT-5.1, etc.).
  • pricing — per-token in/out (when known).
  • contextWindow — max tokens.
  • capabilities — chat, tools, vision, reasoning, json, embeddings.
  • reasoningSupport — explicit reasoning-effort tier when supported.

The catalog drives:

  • The dashboard's model picker (filtered by capability).
  • Cost estimation in pipelines (per-run cost reporting).
  • Model validation (does the configured model actually support the capability the pipeline needs?).

The catalog is sourced from a static JSON snapshot embedded in the agent package, periodically refreshed via a separate task. Plugins can override or extend entries by declaring them in their manifest's models array.

6. Structured Output (askJson)​

askJson<T> is the canonical way to get a typed object back from an LLM:

const { result, model, usage } = await aiFacade.askJson(
	{
		schema: z.object({ name: z.string(), tags: z.array(z.string()) }),
		prompt: 'Extract metadata from: ' + html,
		complexity: 'medium'
	},
	{ userId, workId }
);

The facade:

  1. Converts the Zod schema to a JSON Schema via zod-to-json-schema.
  2. Calls the plugin's askJson method with the JSON Schema.
  3. The plugin uses provider-native structured output if available (OpenAI response_format: json_schema, Anthropic tool use, Gemini responseSchema).
  4. Otherwise it falls back to system-prompt instructions + jsonrepair (best-effort parser for malformed JSON from older models).
  5. The facade re-validates the result against the original Zod schema before returning.

If the result fails Zod validation, the facade retries up to 2 times with the validation error fed back into the prompt — this catches "forgot a field" and "wrong type" errors that fixed-prompt variants struggle with.

7. Streaming Chat​

streamChatCompletion returns an async iterable of ChatCompletionChunks, each containing a delta. The facade handles:

  • Usage accumulation — final chunk emits a usage block with inputTokens / outputTokens / cost (cost computed from the model catalog if available).
  • Error normalisation — provider-specific errors are wrapped in AiFacadeError with operation: 'stream', provider, cause.
  • Cancellation — the consumer can break out of the iterable; the facade signals the underlying plugin to abort.

This is what powers the AI Conversation feature — the API streams NDJSON on top of this iterable.

8. Reasoning & Routing Overrides​

AiRoutingOptions (passed via opts.routing) carries:

FieldEffect
reasoningEffort'low' | 'medium' | 'high'. Picks reasoning-tier model + sets effort param.
temperatureOverride the plugin's default temperature.
maxTokensOverride the plugin's default max tokens.
toolsTool definitions (for tool-calling models).
forceModelBypass tier resolution, use this exact model.
responseFormat'text' | 'json_object' | 'json_schema'.

reasoning.utils.ts (in @ever-works/plugin/ai) holds the canonical mapping between the platform's reasoning levels and provider-native parameters. Adding reasoning support for a new model is a single map entry — the facade and plugins both consult it.

9. Caching​

AiFacadeService.CACHE_TTL = 3_600_000 (1 hour). The facade caches:

  • The resolved provider plugin per (userId, workId) tuple.
  • The resolved provider config per (userId, workId, providerId).
  • The model catalog (process-wide, 1 h).

The cache invalidates automatically on onSettingsUpdated events that match the plugin id or capability. Cache keys are not the place to debug stale settings — use PluginSettingsService.resolveWithSources (see Settings System spec) to see what each tier returned.

10. Validation​

validateConnection(providerId) is the platform's "are these creds and this model wired up correctly?" call. It:

  1. Resolves the plugin and its settings (errors if not enabled).
  2. For each tier (simpleModel, mediumModel, complexModel, defaultModel):
    • Fires a tiny chatCompletion ("ping" prompt).
    • Records success, responseTime, optional error.
  3. Returns a ConnectionValidationResult with overall success + per-tier modelResults.

The dashboard renders this as the test-button output on the plugin settings page. Pipelines call it as a pre-flight check before kicking off long generations — 'failed: anthropic.complexModel claude-sonnet-4-5 returned 401' is much friendlier than a 30-minute generation that dies halfway through.

11. Embeddings​

embeddings(opts) is the same cascade for embedding-capable plugins. Used by the comparison generator and the keyword extractor. The facade returns Float32Array[] to keep the bus small; dimensions vary by model.

12. Error Hierarchy​

All facade errors inherit from AiFacadeError extends FacadeError:

ClassWhen
AiFacadeErrorBase class. Carries operation, provider?, cause?.
Wrapped provider errorsPlugin throws → wrapped with operation context.

Callers can branch on error.operation ('chatCompletion', 'askJson', 'stream', 'validate') and error.provider to surface targeted recovery hints.

13. Performance & Cost​

  • Cache hit path adds < 5 ms. Cold path adds DB roundtrip + plugin instantiation.
  • Streaming has no extra buffering — chunks flow straight through.
  • Cost is reported in the final usage block when the model catalog has pricing for the model. Pipelines aggregate per-run cost into the activity log changelog.

14. Constitution Reconciliation​

PrincipleHow the AI facade respects it
I — Plugin-firstAll AI calls flow through plugin contracts.
II — Capability-drivenResolution by capability, never plugin id.
III — Source-of-truth reposAI facade is in-memory; no data persistence.
IV — Trigger.devHeavy generation runs as background jobs that use the facade.
V — Forward-only migrationsNo DB.
VI — TestsFacade + each AI provider plugin has full Vitest / Jest coverage with mock LangChain.
VII — Secret hygieneAPI keys flow via PluginSettings; never logged; never serialised in errors.
VIII — Plugin countsThe catalog reports counts dynamically.
IX — Behaviour-firstThis spec describes observable behaviour.
X — Backwards-compatIAiFacade is versioned with the SDK.

15. References​