@juspay/neurolink 12.25.5 → 12.26.0

package/dist/providers/systemOneDecision.js

@@ -0,0 +1,443 @@
+import { BaseProvider } from "../core/baseProvider.js";
+import { PROVIDER_DESCRIPTORS_BY_NAME } from "../factories/providerDescriptors.js";
+import { isNeuroLink } from "../neurolink.js";
+import { createProxyFetch } from "../proxy/proxyFetch.js";
+import { ProviderError } from "../types/index.js";
+import { logger } from "../utils/logger.js";
+import { estimateTokens, serializeForEstimate, } from "../utils/tokenEstimation.js";
+/**
+ * Generous enough for a cold start (measured at 2.0–2.7s after idle on Jev)
+ * while still bounded. Every internal caller is fail-open, so this only bites
+ * on a genuine stall.
+ */
+export const DEFAULT_DECISION_TIMEOUT_MS = 5000;
+const DEFAULT_MAX_RETRIES = 1;
+export const isRecord = (v) => typeof v === "object" && v !== null && !Array.isArray(v);
+export const asNumber = (v) => typeof v === "number" && Number.isFinite(v) ? v : undefined;
+const asNumberMap = (v) => {
+    if (!isRecord(v)) {
+        return undefined;
+    }
+    const out = {};
+    for (const [k, raw] of Object.entries(v)) {
+        const n = asNumber(raw);
+        if (n === undefined) {
+            return undefined;
+        }
+        out[k] = n;
+    }
+    return out;
+};
+const asStringMap = (v) => {
+    if (!isRecord(v)) {
+        return undefined;
+    }
+    const out = {};
+    for (const [k, raw] of Object.entries(v)) {
+        if (typeof raw !== "string") {
+            return undefined;
+        }
+        out[k] = raw;
+    }
+    return out;
+};
+/**
+ * Last-resort confidence, when no transport reported one.
+ *
+ * The peak probability is what a calibrated confidence approximates, but it is
+ * NOT the same number, so a threshold tuned against a vendor's calibrated
+ * figure does not transfer unexamined; an even distribution lands near 1/N
+ * rather than 0.
+ */
+function deriveConfidence(probabilities) {
+    const values = Object.values(probabilities);
+    return values.length > 0 ? Math.max(...values) : 0;
+}
+/**
+ * Validate one answer off the wire. Returns null rather than throwing so a
+ * single malformed answer degrades to "unanswered" instead of failing the
+ * whole batch — the batch may hold hundreds of usable answers.
+ */
+function parseDecisionAnswer(raw, reportedConfidence) {
+    if (!isRecord(raw)) {
+        return null;
+    }
+    switch (raw.type) {
+        // "noul" is the System One wire's own spelling; "boolean" is the neutral
+        // one (Vercel's gateway), which renames both the type and the field.
+        // Accepting both keeps one parser for every transport.
+        case "noul":
+        case "boolean": {
+            const probability = asNumber(raw.noul) ?? asNumber(raw.probability);
+            return probability === undefined
+                ? null
+                : { type: "boolean", probability };
+        }
+        case "choice": {
+            const probabilities = asNumberMap(raw.probabilities);
+            if (typeof raw.choice !== "string" || !probabilities) {
+                return null;
+            }
+            return {
+                type: "choice",
+                choice: raw.choice,
+                confidence: asNumber(raw.confidence) ??
+                    reportedConfidence ??
+                    deriveConfidence(probabilities),
+                probabilities,
+            };
+        }
+        case "score": {
+            const score = asNumber(raw.score);
+            const legend = asStringMap(raw.legend);
+            const probabilities = asNumberMap(raw.probabilities);
+            if (score === undefined || !legend || !probabilities) {
+                return null;
+            }
+            return {
+                type: "score",
+                score,
+                confidence: asNumber(raw.confidence) ??
+                    reportedConfidence ??
+                    deriveConfidence(probabilities),
+                legend,
+                probabilities,
+            };
+        }
+        default:
+            return null;
+    }
+}
+/**
+ * FastAPI's request-validation envelope, `{"detail":[{loc, msg, input}]}`,
+ * reduced to its field paths and messages. `input` echoes the caller's own
+ * request back — possibly the whole state — so it is never surfaced.
+ */
+export function describeValidationErrors(detail) {
+    const fields = detail
+        .filter(isRecord)
+        .map((d) => {
+        const loc = Array.isArray(d.loc)
+            ? d.loc
+                .filter((p) => typeof p === "string" || typeof p === "number")
+                .join(".")
+            : "";
+        const msg = typeof d.msg === "string" ? d.msg : "invalid";
+        return loc ? `${loc}: ${msg}` : msg;
+    })
+        .slice(0, 5);
+    return fields.join("; ") || "unspecified field";
+}
+const sleep = (ms) => new Promise((r) => setTimeout(r, ms));
+/**
+ * `estimateTokens` assumes ~4 characters per token. That holds for English
+ * and is several times too generous for other scripts on an English
+ * tokenizer, so when a provider declares a rate, non-ASCII characters are
+ * counted at it instead. ASCII text is estimated exactly as before.
+ */
+function estimateDecisionStateTokens(text, nonAsciiTokensPerChar) {
+    if (nonAsciiTokensPerChar === undefined) {
+        return estimateTokens(text);
+    }
+    const chars = [...text];
+    const isAscii = (c) => (c.codePointAt(0) ?? 0) <= 0x7f;
+    const ascii = chars.filter(isAscii).join("");
+    const nonAsciiCount = chars.length - ascii.length;
+    return (estimateTokens(ascii) + Math.ceil(nonAsciiCount * nonAsciiTokensPerChar));
+}
+/**
+ * The shared half of every "System One" decision provider — a model that takes
+ * one `state` plus typed questions and returns one typed answer each, with no
+ * text anywhere.
+ *
+ * Vendors differ in where the request goes, how it is authenticated, what the
+ * body carries and what their error envelopes look like. They do not differ in
+ * anything else, so everything else lives here: the missing-key check, the
+ * auth circuit breaker, the timeout and retry loop, answer parsing and usage.
+ * A new decision provider implements the abstract hooks and nothing more.
+ */
+export class SystemOneDecisionProvider extends BaseProvider {
+    proxyFetch;
+    maxRetries;
+    /**
+     * Set once the key has been rejected. A bad key cannot fix itself, and
+     * without this every later request would pay a full round trip (~265ms
+     * measured on Jev) to be told so again — a permanent tax on the hot path for
+     * a misconfiguration.
+     */
+    authFailed = false;
+    constructor(modelName, providerName, sdk) {
+        super(modelName, providerName, isNeuroLink(sdk) ? sdk : undefined);
+        this.proxyFetch = createProxyFetch();
+        this.maxRetries = DEFAULT_MAX_RETRIES;
+    }
+    /**
+     * Why this instance cannot send a request even though it has a key, such as
+     * a provider with no built-in endpoint and none configured. Undefined when
+     * nothing is missing.
+     */
+    missingConfigMessage() {
+        return undefined;
+    }
+    /**
+     * The System One wire calls the yes/no primitive `noul`; this codebase and
+     * every SDK exposing it call it `boolean`.
+     */
+    encodeQuestion(question) {
+        if (question.type === "boolean") {
+            return {
+                type: "noul",
+                instructions: question.instructions,
+                ...(question.criteria ? { criteria: question.criteria } : {}),
+            };
+        }
+        return question;
+    }
+    /** Per-question confidence a transport reports outside the answer objects. */
+    reportedConfidence(_payload) {
+        return {};
+    }
+    resolveResponseModel(payload, requestedModel) {
+        return typeof payload.model === "string" ? payload.model : requestedModel;
+    }
+    /** Extra fields for the credential-rejected warning. */
+    diagnosticContext() {
+        return {};
+    }
+    // ===== BaseProvider obligations a text-less model satisfies identically =====
+    getProviderName() {
+        return this.providerName;
+    }
+    getDefaultDecisionModel() {
+        return this.getDefaultModel();
+    }
+    supportsTools() {
+        return false;
+    }
+    /**
+     * A decision model produces no text, so there is no chat model behind it.
+     * Unreachable while the descriptor declares `inferenceKinds: ["decide"]`;
+     * kept actionable for the case where something routes here anyway.
+     */
+    getAISDKModel() {
+        throw new ProviderError(`${this.vendorDisplayName()} is a decision-only provider; it generates no text, so chat completions are not available. Use \`decide()\` instead, or pick a different provider for \`generate()\` / \`stream()\`.`, this.providerName);
+    }
+    async executeStream(_options, _analysisSchema) {
+        throw new ProviderError(`${this.vendorDisplayName()} is a decision-only provider; streaming is not available. Use \`decide()\`, or pick another provider for \`stream()\`.`, this.providerName);
+    }
+    /**
+     * Must RETURN the error, never throw it (CLAUDE.md rule 6).
+     */
+    formatProviderError(error) {
+        if (error instanceof ProviderError) {
+            return error;
+        }
+        const message = error instanceof Error ? error.message : String(error);
+        return new ProviderError(`${this.vendorLabel()} request failed: ${message}`, this.providerName);
+    }
+    // ===== The decide inference type =====
+    /**
+     * Evaluate a state against a batch of typed questions.
+     *
+     * Throws a {@link ProviderError} carrying a normalised {@link DecisionError}
+     * in `cause`. Internal hot paths should prefer `NeuroLink.tryDecide()`,
+     * which returns null instead.
+     */
+    async decide(request) {
+        const label = this.vendorLabel();
+        if (!this.decisionApiKey()) {
+            throw this.decisionError({
+                kind: "authentication",
+                message: this.missingKeyMessage(),
+                retryable: false,
+            });
+        }
+        const missingConfig = this.missingConfigMessage();
+        if (missingConfig) {
+            throw this.decisionError({
+                kind: "invalid_request",
+                message: missingConfig,
+                retryable: false,
+            });
+        }
+        if (this.authFailed) {
+            throw this.decisionError({
+                kind: "authentication",
+                message: `${label} rejected this API key earlier; not retrying. Fix the key and construct a new provider.`,
+                retryable: false,
+            });
+        }
+        const questionEntries = Object.entries(request.questions);
+        if (questionEntries.length === 0) {
+            throw this.decisionError({
+                kind: "invalid_request",
+                message: "At least one question is required.",
+                retryable: false,
+            });
+        }
+        const resolvedModel = request.model ?? this.modelName ?? this.getDefaultModel();
+        this.assertWithinDecisionLimits(request, resolvedModel, questionEntries.length);
+        const wireQuestions = Object.fromEntries(questionEntries.map(([id, question]) => [
+            id,
+            this.encodeQuestion(question),
+        ]));
+        const body = JSON.stringify(this.buildDecisionBody(request.state, wireQuestions, resolvedModel));
+        const timeoutMs = request.timeoutMs ??
+            this.getDescriptorDecideMs() ??
+            this.defaultTimeout ??
+            DEFAULT_DECISION_TIMEOUT_MS;
+        let lastError;
+        for (let attempt = 0; attempt <= this.maxRetries; attempt++) {
+            const startedAt = Date.now();
+            try {
+                const timeout = AbortSignal.timeout(timeoutMs);
+                const signal = request.signal
+                    ? AbortSignal.any([request.signal, timeout])
+                    : timeout;
+                const response = await this.proxyFetch(this.decisionEndpoint(), {
+                    method: "POST",
+                    headers: this.decisionHeaders(),
+                    body,
+                    signal,
+                });
+                const latencyMs = Date.now() - startedAt;
+                const requestId = this.readRequestId(response.headers);
+                const upstreamMs = Number(response.headers.get("x-envoy-upstream-service-time")) ||
+                    undefined;
+                const payload = await response.json().catch(() => null);
+                if (!response.ok) {
+                    lastError = this.parseDecisionError(response.status, payload, requestId);
+                    if (lastError.kind === "authentication") {
+                        this.authFailed = true;
+                        // The breaker is right to trip — none of these clear themselves
+                        // inside one process — but it must not assert WHICH credential
+                        // problem it was. A gateway 403 with a valid key and no card on
+                        // file is not a rejected key, and saying so sends the reader off
+                        // to regenerate the one thing that works.
+                        logger.warn(`${label}: credential not accepted — disabling this provider instance. ${lastError.message}`, {
+                            status: lastError.status,
+                            ...this.diagnosticContext(),
+                            requestId,
+                        });
+                    }
+                    if (!lastError.retryable || attempt === this.maxRetries) {
+                        throw this.decisionError(lastError);
+                    }
+                    await sleep(2 ** attempt * 250 + Math.random() * 250);
+                    continue;
+                }
+                if (!isRecord(payload) || !isRecord(payload.answers)) {
+                    throw this.decisionError({
+                        kind: "server",
+                        message: `${label} returned a response without an answers map.`,
+                        status: response.status,
+                        requestId,
+                        retryable: false,
+                    });
+                }
+                const reported = this.reportedConfidence(payload);
+                const answers = {};
+                for (const [id, raw] of Object.entries(payload.answers)) {
+                    const parsed = parseDecisionAnswer(raw, reported[id]);
+                    if (parsed) {
+                        answers[id] = parsed;
+                    }
+                    else {
+                        logger.warn(`${label}: dropped unparseable answer "${id}"`, {
+                            requestId,
+                        });
+                    }
+                }
+                const usage = isRecord(payload.usage) ? payload.usage : {};
+                return {
+                    // `resolvedModel`, not `this.modelName`: when the caller pinned a
+                    // model for this one request and the response omits its own, the
+                    // instance default would be reported instead of the model actually
+                    // asked for.
+                    model: this.resolveResponseModel(payload, resolvedModel),
+                    provider: this.providerName,
+                    answers,
+                    // Two spellings for one field. The System One wire sends
+                    // `input_tokens`; Vercel's gateway sends `inputTokens`. Reading only
+                    // the first priced every gateway decision at exactly $0, silently.
+                    usage: {
+                        inputTokens: asNumber(usage.input_tokens) ?? asNumber(usage.inputTokens) ?? 0,
+                        outputTokens: asNumber(usage.output_tokens) ??
+                            asNumber(usage.outputTokens) ??
+                            0,
+                    },
+                    requestId,
+                    latencyMs,
+                    upstreamMs,
+                };
+            }
+            catch (error) {
+                if (error instanceof ProviderError) {
+                    throw error;
+                }
+                // An abort from the caller's own signal must not be retried — the
+                // turn is already being torn down.
+                const aborted = request.signal?.aborted === true;
+                const isTimeout = error instanceof Error && error.name === "TimeoutError";
+                lastError = {
+                    kind: aborted ? "network" : isTimeout ? "timeout" : "network",
+                    message: error instanceof Error ? error.message : String(error),
+                    retryable: !aborted,
+                };
+                if (aborted || attempt === this.maxRetries) {
+                    throw this.decisionError(lastError);
+                }
+                await sleep(2 ** attempt * 250 + Math.random() * 250);
+            }
+        }
+        throw this.decisionError(lastError ?? {
+            kind: "network",
+            message: `${label} request failed.`,
+            retryable: false,
+        });
+    }
+    /**
+     * Refuse a request the model cannot read in full. An encoder cuts the state
+     * off past its window without saying so, so without this a decision would
+     * be made on input the model never saw — and reported as if it had.
+     */
+    assertWithinDecisionLimits(request, model, questionCount) {
+        const limits = PROVIDER_DESCRIPTORS_BY_NAME.get(this.providerName)?.decisionLimits;
+        if (!limits) {
+            return;
+        }
+        const label = this.vendorLabel();
+        if (questionCount > limits.maxQuestions) {
+            throw this.decisionError({
+                kind: "max_tokens_exceeded",
+                message: `${label} accepts at most ${limits.maxQuestions} questions per request; this one has ${questionCount}. Ask fewer, or use a decision provider without this cap.`,
+                retryable: false,
+            });
+        }
+        const modelLimits = limits.models?.[model];
+        const maxStateTokens = modelLimits?.maxStateTokens ?? limits.maxStateTokens;
+        const stateTokens = estimateDecisionStateTokens(serializeForEstimate(request.state), modelLimits?.nonAsciiTokensPerChar ?? limits.nonAsciiTokensPerChar);
+        if (stateTokens > maxStateTokens) {
+            throw this.decisionError({
+                kind: "max_tokens_exceeded",
+                message: `The state is ~${stateTokens} tokens; ${label}'s "${model}" model reads at most ${maxStateTokens}. Shorten the state, or use a decision provider with a larger window.`,
+                retryable: false,
+            });
+        }
+    }
+    /**
+     * Wrap a normalised decision failure in the provider error type, keeping
+     * the structured detail reachable via `cause` for callers that branch on
+     * `kind` (authentication vs rate_limit vs max_tokens_exceeded).
+     */
+    decisionError(detail) {
+        const error = new ProviderError(detail.message, this.providerName);
+        Object.defineProperty(error, "cause", {
+            value: detail,
+            enumerable: false,
+            writable: true,
+            configurable: true,
+        });
+        return error;
+    }
+}

package/dist/providers/typesafe.d.ts

@@ -1,6 +1,5 @@
-import type { AIProviderName } from "../constants/enums.js";
-import { BaseProvider } from "../core/baseProvider.js";
-import type { DecisionModelCard, DecisionRequest, DecisionResult, LanguageModel, NeurolinkCredentials, StreamOptions, StreamResult, ValidationSchema } from "../types/index.js";
+import type { DecisionError, DecisionModelCard, DecisionQuestion, DecisionState, NeurolinkCredentials } from "../types/index.js";
+import { SystemOneDecisionProvider } from "./systemOneDecision.js";
 /**
  * The binding limit in practice: `state` plus the SINGLE LONGEST question.
  * Measured by bisection at single-character resolution — 33,002 reported
@@ -19,13 +18,12 @@ export declare const TYPESAFE_MAX_REQUEST_TOKENS = 64000;
  *
  * Jev is a "System One" model: it takes one `state` plus a map of typed
  * questions and returns one typed answer each, evaluated in a single
- * parallel pass. It emits no text, so `generate()` / `stream()` are not
- * available and `getAISDKModel()` throws — the same shape Voyage and Jina
- * already use for embedding-only providers.
+ * parallel pass. The request loop, retries, auth circuit breaker and answer
+ * parsing are shared with every decision provider in
+ * {@link SystemOneDecisionProvider}; this class holds only what is TypeSafe's.
  *
  * Its descriptor declares `inferenceKinds: ["decide"]`, which keeps it out
- * of auto-select and the health sweep, so those throws should be unreachable
- * in practice.
+ * of auto-select and the health sweep.
  *
  * Two measured properties shape the API: latency is flat in question count
  * (1 question 393ms, 400 questions 465ms) while concurrent requests queue.
@@ -35,54 +33,47 @@ export declare const TYPESAFE_MAX_REQUEST_TOKENS = 64000;
  * billed against a TypeSafe key. The gateway one is Vercel's AI Gateway
  * evaluation-model route, billed against a Vercel account — useful to a host
  * that already has one and does not want a second vendor relationship. They
- * differ on the wire (vocabulary, where the model is named, and whether
- * `confidence` is reported at all) and not at all in `DecisionResult`; the
- * one thing to know is that gateway confidence is DERIVED from the
- * probability distribution rather than being the vendor's calibrated figure,
- * so a threshold tuned on one transport should be re-checked on the other.
+ * differ on the wire (vocabulary, where the model is named, and where
+ * `confidence` is reported) and not at all in `DecisionResult`.
  *
  * @see https://docs.typesafe.ai/api
  */
-export declare class TypeSafeProvider extends BaseProvider {
+export declare class TypeSafeProvider extends SystemOneDecisionProvider {
     private readonly apiKey;
     private readonly baseURL;
-    private readonly proxyFetch;
-    private readonly maxRetries;
+    private readonly gatewayURL;
     /** Which wire this instance speaks. Resolved once, in the constructor. */
     private readonly transport;
-    /**
-     * Set once the key has been rejected. A bad key cannot fix itself, and
-     * without this every later request would pay a full round trip (~265ms
-     * measured) to be told so again — a permanent tax on the hot path for a
-     * misconfiguration.
-     */
-    private authFailed;
     constructor(modelName?: string, sdk?: unknown, _region?: string, credentials?: NeurolinkCredentials["typesafe"]);
-    protected getProviderName(): AIProviderName;
     protected getDefaultModel(): string;
-    protected getDefaultDecisionModel(): string | undefined;
-    supportsTools(): boolean;
+    protected vendorLabel(): string;
+    protected vendorDisplayName(): string;
+    protected decisionApiKey(): string;
+    protected missingKeyMessage(): string;
     /**
-     * Jev produces no text, so there is no chat model behind this provider.
-     * Unreachable while the descriptor declares `inferenceKinds: ["decide"]`;
-     * kept actionable for the case where something routes here anyway.
+     * The gateway already speaks the neutral vocabulary this codebase uses, so
+     * only the direct transport needs the `noul` rename.
      */
-    protected getAISDKModel(): LanguageModel;
-    protected executeStream(_options: StreamOptions, _analysisSchema?: ValidationSchema): Promise<StreamResult>;
+    protected encodeQuestion(question: DecisionQuestion): Record<string, unknown>;
     /**
-     * Must RETURN the error, never throw it (CLAUDE.md rule 6).
+     * The gateway names the model in a header and rejects a body that carries
+     * one; the direct API requires it in the body.
      */
-    protected formatProviderError(error: unknown): Error;
+    protected buildDecisionBody(state: DecisionState, questions: Record<string, Record<string, unknown>>, model: string): Record<string, unknown>;
+    protected parseDecisionError(status: number, payload: unknown, requestId: string | undefined): DecisionError;
     /**
-     * Evaluate a state against a batch of typed questions.
-     *
-     * Throws a {@link ProviderError} carrying a normalised {@link DecisionError}
-     * in `cause`. Internal hot paths should prefer `NeuroLink.tryDecide()`,
-     * which returns null instead.
+     * The gateway does not omit confidence after all — it RELOCATES it, to
+     * `providerMetadata.typesafe.confidence.<questionId>`, and leaves the answer
+     * objects themselves without one. Preferring it over the derived peak
+     * matters because the two disagree on any distribution that is not already
+     * near-certain, and the router's asymmetric bars are tuned against the
+     * vendor's calibrated figure, not against max(probabilities).
      */
-    decide(request: DecisionRequest): Promise<DecisionResult>;
+    protected reportedConfidence(payload: unknown): Record<string, number>;
+    protected readRequestId(headers: Headers): string | undefined;
+    protected diagnosticContext(): Record<string, unknown>;
     /** Where a decision request goes for this instance's transport. */
-    private endpoint;
+    protected decisionEndpoint(): string;
     /**
      * Headers for this instance's transport.
      *
@@ -92,7 +83,7 @@ export declare class TypeSafeProvider extends BaseProvider {
      * carries no `model`), and `ai-evaluation-model-specification-version`
      * selects the request/response shape this provider parses.
      */
-    private headers;
+    protected decisionHeaders(): Record<string, string>;
     /** Which transport this instance resolved to. Exposed for diagnostics. */
     getDecisionTransport(): "direct" | "gateway";
     /**
@@ -103,10 +94,4 @@ export declare class TypeSafeProvider extends BaseProvider {
      * this key.
      */
     listDecisionModels(): Promise<DecisionModelCard[]>;
-    /**
-     * Wrap a normalised decision failure in the provider error type, keeping
-     * the structured detail reachable via `cause` for callers that branch on
-     * `kind` (authentication vs rate_limit vs max_tokens_exceeded).
-     */
-    private decisionError;
 }