@plurnk/plurnk-providers 1.16.4 → 1.17.0

package/src/providerError.ts

@@ -7,6 +7,7 @@ export type ProviderErrorKind =
     | "network_failure"
     | "deadline_exceeded"
     | "model_refused"
+    | "request_rejected"
     | "invalid_response"
     | "unauthorized"
     | "quota_exceeded"
@@ -16,6 +17,7 @@ export type ProviderErrorKind =
 
 const defaultStatus = (kind: ProviderErrorKind): number => {
     switch (kind) {
+        case "request_rejected": return 400;
         case "unauthorized": return 401;
         case "quota_exceeded": return 402;
         case "capacity_exceeded": return 413;
@@ -36,6 +38,7 @@ const retryable = (kind: ProviderErrorKind): boolean => {
             return true;
         case "deadline_exceeded":
         case "invalid_response":
+        case "request_rejected":
         case "grammar_invalid":
         case "capacity_exceeded":
         case "resource_interrupted":
@@ -60,6 +63,7 @@ const buildProblem = (
         deadline_exceeded: "deadline-exceeded",
         model_refused: "model-refused",
         invalid_response: "invalid-response",
+        request_rejected: "request-rejected",
         unauthorized: "unauthorized",
         quota_exceeded: "quota-exceeded",
         grammar_invalid: "grammar-invalid",

package/src/sdkModels.test.ts

@@ -1,6 +1,6 @@
 import test from "node:test";
 import { strict as assert } from "node:assert";
-import { configuredProviderInfo, createEmbeddingModel, createSdkModel, providerReadiness } from "./sdkModels.ts";
+import { configuredProviderInfo, createSdkModel, providerReadiness } from "./sdkModels.ts";
 
 test("{§provider-fact-authority} one env declaration holds one credential name", () => {
     assert.deepEqual(configuredProviderInfo("acme-cloud", {
@@ -301,44 +301,3 @@ test("createSdkModel fails clearly for a declared but unsupported SDK package",
         /Models.dev declares unsupported AI SDK package @acme\/ai-sdk/,
     );
 });
-
-test("{§provider-embedding-resolution} compatible catalog and declared routes share the standard embedding adapter", () => {
-    const cloudflare = createEmbeddingModel("cloudflare-workers-ai", "@cf/qwen/qwen3-embedding-0.6b", {
-        CLOUDFLARE_ACCOUNT_ID: "account",
-        CLOUDFLARE_API_KEY: "key",
-    });
-    assert.equal(cloudflare?.providerId, "cloudflare-workers-ai");
-    assert.equal(cloudflare?.embeddingModel.provider, "cloudflare-workers-ai.embedding");
-    assert.equal(cloudflare?.embeddingModel.modelId, "@cf/qwen/qwen3-embedding-0.6b");
-
-    const fireworks = createEmbeddingModel("fireworks-ai", "fireworks/qwen3-embedding-8b", {
-        FIREWORKS_API_KEY: "key",
-    });
-    assert.equal(fireworks?.providerId, "fireworks-ai");
-    assert.equal(fireworks?.embeddingModel.provider, "fireworks-ai.embedding");
-
-    const local = createEmbeddingModel("local-embed", "Qwen/Qwen3-Embedding-0.6B", {
-        PLURNK_PROVIDERS_PROVIDER_LOCAL_EMBED_NPM: "@ai-sdk/openai-compatible",
-        PLURNK_PROVIDERS_PROVIDER_LOCAL_EMBED_BASE_URL: "http://127.0.0.1:8080/v1",
-    });
-    assert.equal(local?.providerId, "local-embed");
-    assert.equal(local?.embeddingModel.provider, "local-embed.embedding");
-});
-
-test("{§provider-embedding-resolution} OpenRouter uses its official embedding model", () => {
-    const resolved = createEmbeddingModel("openrouter", "openai/text-embedding-3-small", {
-        OPENROUTER_API_KEY: "key",
-        OPENROUTER_HTTP_REFERER: "https://github.com/plurnk/plurnk-service",
-        OPENROUTER_APP_TITLE: "Plurnk",
-    });
-    assert.equal(resolved?.providerId, "openrouter");
-    assert.equal(resolved?.embeddingModel.provider, "openrouter");
-    assert.equal(resolved?.embeddingModel.modelId, "openai/text-embedding-3-small");
-});
-
-test("{§provider-embedding-resolution} a generation-only SDK fails before transport", () => {
-    assert.throws(
-        () => createEmbeddingModel("cerebras", "text-embedding", { CEREBRAS_API_KEY: "key" }),
-        /@ai-sdk\/cerebras does not expose an embedding model/,
-    );
-});

package/src/sdkModels.ts

@@ -6,15 +6,12 @@ import { createGoogle } from "@ai-sdk/google";
 import { createGroq } from "@ai-sdk/groq";
 import { createMistral } from "@ai-sdk/mistral";
 import { createOpenAI } from "@ai-sdk/openai";
-import { createOpenAICompatible } from "@ai-sdk/openai-compatible";
 import { createTogetherAI } from "@ai-sdk/togetherai";
 import { createXai } from "@ai-sdk/xai";
-import type { EmbeddingModelV4, EmbeddingModelV4Result } from "@ai-sdk/provider";
 import { createOpenRouter, type OpenRouterChatSettings } from "@openrouter/ai-sdk-provider";
 import {
     resolveModel,
     isProviderCredentialName,
-    lookup,
     lookupProvider,
     providerCatalogSnapshot,
     type ProviderInfo,
@@ -23,13 +20,10 @@ import {
     Validator,
     type ModelReadiness,
     type ModelReadinessCause,
-    type ProviderRequestAccounting,
-    type ProviderUsage,
     type ReasoningPolicy,
 } from "@plurnk/plurnk-contracts";
 import type { LanguageModel } from "ai";
-import { providerCostNormalizer, validateProviderRequestAccounting } from "./accounting.ts";
-import { estimateProviderCost, resolveProviderCost } from "./cost.ts";
+import { providerCostNormalizer } from "./accounting.ts";
 import type { AiSdkProviderOptions, CacheAffinity } from "./AiSdkProvider.ts";
 import type { ProviderCostNormalizer } from "./types.ts";
 
@@ -47,17 +41,6 @@ export type SdkModel = {
     readonly catalog: ProviderInfo | null;
 };
 
-export type EmbeddingModelResolution = {
-    readonly embeddingModel: EmbeddingModelV4;
-    readonly providerId: string;
-    readonly modelId: string;
-    readonly normalizeAccounting: (evidence: {
-        readonly outcome: "response" | "error";
-        readonly status?: number;
-        readonly result?: EmbeddingModelV4Result;
-    }) => ProviderRequestAccounting;
-};
-
 const cacheControl = { type: "ephemeral" as const };
 // The release generator admits only packages implemented by this package.
 // Derive runtime readiness from that same pinned provider projection instead
@@ -485,108 +468,3 @@ export const createSdkModel = (
             throw new Error(`${provider} provider: Models.dev declares unsupported AI SDK package ${catalog.npm}`);
     }
 };
-
-// {§provider-embedding-resolution} Provider construction owns transport and
-// authentication only. Model-space facts stay in the embedding profile.
-export const createEmbeddingModel = (
-    provider: string,
-    model: string,
-    env: NodeJS.ProcessEnv,
-    baseUrlOverride?: string,
-): EmbeddingModelResolution | null => {
-    const resolved = resolveSdkProvider(provider, env, baseUrlOverride);
-    if (resolved === null) return null;
-    const { catalog, url } = resolved;
-    const directCost = providerCostNormalizer(catalog.npm);
-    const modelCost = lookup(provider, model)?.cost ?? lookup(catalog.id, model)?.cost;
-    const rates = modelCost === undefined
-        ? null
-        : {
-            input: modelCost.inputPer1M,
-            output: modelCost.outputPer1M,
-            ...(modelCost.reasoningPer1M === undefined ? {} : { reasoning: modelCost.reasoningPer1M }),
-            ...(modelCost.cacheReadPer1M === undefined ? {} : { cacheRead: modelCost.cacheReadPer1M }),
-            ...(modelCost.cacheWritePer1M === undefined ? {} : { cacheWrite: modelCost.cacheWritePer1M }),
-        };
-    const normalizeAccounting: EmbeddingModelResolution["normalizeAccounting"] = (evidence) => {
-        const tokens = evidence.result?.usage?.tokens;
-        const usage: ProviderUsage | undefined = Number.isSafeInteger(tokens) && tokens! >= 0
-            ? { inputTokens: tokens!, outputTokens: 0, totalTokens: tokens! }
-            : undefined;
-        const body = evidence.result?.response?.body;
-        const rawUsage = typeof body === "object" && body !== null && !Array.isArray(body)
-            ? (body as Record<string, unknown>).usage
-            : undefined;
-        const direct = directCost?.({
-            providerMetadata: evidence.result?.providerMetadata,
-            usage: rawUsage,
-            response: { headers: evidence.result?.response?.headers },
-        });
-        const cost = resolveProviderCost(
-            direct,
-            estimateProviderCost(
-                usage,
-                rates,
-                `Models.dev ${catalog.id}/${model} rates`,
-            ),
-        );
-        return validateProviderRequestAccounting({
-            provider: catalog.id,
-            model,
-            outcome: evidence.outcome,
-            ...(evidence.status === undefined ? {} : { status: evidence.status }),
-            ...(usage === undefined ? {} : { usage }),
-            cost,
-        });
-    };
-    const result = (embeddingModel: EmbeddingModelV4): EmbeddingModelResolution => ({
-        embeddingModel,
-        providerId: catalog.id,
-        modelId: model,
-        normalizeAccounting,
-    });
-
-    switch (catalog.npm) {
-        case "@ai-sdk/openai":
-            return result(createOpenAI({ apiKey: requireApiKey(provider, env, catalog), baseURL: url }).embedding(model));
-        case "@ai-sdk/google":
-            return result(createGoogle({ apiKey: requireApiKey(provider, env, catalog), baseURL: url }).embedding(model));
-        case "@ai-sdk/mistral":
-            return result(createMistral({ apiKey: requireApiKey(provider, env, catalog), baseURL: url }).embedding(model));
-        case "@ai-sdk/deepinfra":
-            return result(createDeepInfra({ apiKey: requireApiKey(provider, env, catalog), baseURL: url }).embeddingModel(model));
-        case "@ai-sdk/togetherai":
-            return result(createTogetherAI({ apiKey: requireApiKey(provider, env, catalog), baseURL: url }).embeddingModel(model));
-        case "@ai-sdk/amazon-bedrock":
-            return result(createAmazonBedrock({
-                region: env.AWS_REGION ?? env.AWS_DEFAULT_REGION,
-                accessKeyId: env.AWS_ACCESS_KEY_ID,
-                secretAccessKey: env.AWS_SECRET_ACCESS_KEY,
-                sessionToken: env.AWS_SESSION_TOKEN,
-                apiKey: env.AWS_BEARER_TOKEN_BEDROCK,
-                baseURL: url,
-            }).embedding(model));
-        case "@openrouter/ai-sdk-provider":
-            return result(createOpenRouter({
-                apiKey: requireApiKey(provider, env, catalog),
-                baseURL: url,
-                headers: openRouterHeaders(provider, env, catalog),
-            }).textEmbeddingModel(model));
-        case "@ai-sdk/openai-compatible": {
-            if (url === undefined) {
-                throw new Error(`${provider} provider: Models.dev supplies no API URL and no base URL was configured`);
-            }
-            const keyNames = configuredKeyNames(provider, env, catalog);
-            const key = keyNames.length === 0 ? undefined : requireApiKey(provider, env, catalog);
-            return result(createOpenAICompatible({
-                name: provider,
-                baseURL: url,
-                ...(key === undefined ? {} : { apiKey: key }),
-            }).embeddingModel(model));
-        }
-        default:
-            throw new Error(
-                `${provider} provider: ${catalog.npm} does not expose an embedding model through Plurnk's AI SDK adapter`,
-            );
-    }
-};

package/src/types.ts

@@ -187,10 +187,10 @@ export interface ProviderAssistant<TFinish extends ProviderAttemptFinishReason =
     readonly meanLogprob?: number;
 }
 
+// {§provider-grammar-transport} — transport evidence for an operator's grammar: the exact
+// pre-projection response, where the projected content starts within it, and whether the
+// grammar reached the wire. Evidence only; nothing grades it.
 export interface GrammarEvidence {
-    // Exact pre-projection response represented by the provider. A generated
-    // rail's response root composes any template prefix before @plurnk/gbnf
-    // grades it. Offsets are Unicode code points, matching validator verdicts.
     readonly input: string;
     readonly contentStart: number;
     readonly transported: boolean;
@@ -203,7 +203,7 @@ export interface ProviderResponse<TFinish extends ProviderAttemptFinishReason =
     // failover that preceded this response. {§provider-request-accounting}
     readonly accounting: readonly ProviderRequestAccounting[];
     readonly capacity: ProviderRequestCapacity;
-    // {§gbnf-response-observation} — evidence only; the consumer owns the verdict.
+    // {§provider-grammar-transport} — transport evidence; present only when a grammar was configured.
     readonly grammarEvidence?: GrammarEvidence;
     // Per-turn provider→client metadata bag: the backend's non-standard top-level
     // response fields passed through verbatim. Monetary values carry their own
@@ -250,12 +250,10 @@ export interface Provider {
     // Optional package-authored folksonomy evaluated by the consumer immediately
     // before a provider emission attempt ({§plugin-attribution}).
     attributions?(context: PluginAttributionContext): PluginAttribution;
-    // `grammar` is an optional GBNF string (canonically @plurnk/plurnk-contracts'
-    // plurnk.gemma.gbnf or plurnk.qwen.gbnf, possibly root-substituted by the consumer). Backends that
-    // support grammar-constrained sampling attach it verbatim; all others
-    // ignore it. The provider never chooses or modifies the grammar — whether
-    // to constrain and which root variant to send is consumer policy
-    // ({§gbnf-response-observation}).
+    // `grammar` is an optional GBNF string, the operator's own file read by the consumer.
+    // Backends that support grammar-constrained sampling attach it verbatim; all others
+    // ignore it. The provider never chooses or modifies the grammar
+    // ({§provider-grammar-transport}).
     //
     // `maxOutputTokens` may tighten the provider's configured total output
     // budget for this call. It includes visible output and hidden reasoning;