refactor: resolve request defaults once in embedTexts (CL-9052)

Author: TheGreatAxios

src/embed.ts

@@ -1,13 +1,16 @@
 import { type } from "arktype";
 import {
   classifyProtocolMismatch,
+  createDefaultRetryPolicy,
   createDefaultScheduler,
   type BuiltRequest,
 } from "@intx/inference";
 import type { RetryPolicy } from "@intx/types/runtime";
 
 import {
+  DEFAULT_TIMEOUT_MS,
   EmbeddingRequestError,
+  extractRetryAfterMs,
   runJSONRequest,
   type RequestDependencies,
   type RetryAfterExtractor,
@@ -58,10 +61,11 @@ const PROBE_TEXT = "embedding dimension probe";
 export type EmbedOptions = {
   /** Defaults to global `fetch` and Interchange's default scheduler. */
   deps?: RequestDependencies;
+  /** Defaults to Interchange's policy: back off retryables, abort the rest. */
   retryPolicy?: RetryPolicy;
   /**
    * Reads `Retry-After` off a failed response, for a provider that signals
-   * pacing its own way. Named after `ProviderAdapter.extractRetryAfterMs`.
+   * pacing its own way. Defaults to seconds or HTTP-date parsing. Named after `ProviderAdapter.extractRetryAfterMs`.
    * There is no adapter object here because every provider serves the one
    * `/v1/embeddings` shape, so this hangs off the options instead.
    */
@@ -187,13 +191,17 @@ export async function embedTexts(
     scheduler: createDefaultScheduler(),
   };
 
+  const retryPolicy = options.retryPolicy ?? createDefaultRetryPolicy();
+  const timeoutMs = valid.timeoutMs ?? DEFAULT_TIMEOUT_MS;
+  const extractRetryAfter = options.extractRetryAfterMs ?? extractRetryAfterMs;
+
   const vectors: number[][] = [];
   for (const batch of batches(texts, valid.batchSize ?? DEFAULT_BATCH_SIZE)) {
     const request = buildRequest(valid, batch);
     const body = await runJSONRequest(request, deps, {
-      retryPolicy: options.retryPolicy,
-      timeoutMs: valid.timeoutMs,
-      extractRetryAfterMs: options.extractRetryAfterMs,
+      retryPolicy,
+      timeoutMs,
+      extractRetryAfterMs: extractRetryAfter,
       signal: options.signal,
     });
     vectors.push(...parseResponse(body, request.url, batch.length));

src/request.ts

@@ -3,7 +3,6 @@ import {
   classifyHTTPError,
   classifyNetworkError,
   classifyProtocolMismatch,
-  createDefaultRetryPolicy,
   type BuiltRequest,
   type Dependencies,
 } from "@intx/inference";
@@ -54,28 +53,26 @@ export class EmbeddingRequestError extends Error {
 export type RetryAfterExtractor = (headers: Headers) => number | undefined;
 
 type RunRequestOptions = {
-  /** Defaults to Interchange's policy: back off retryables, abort the rest. */
-  retryPolicy: RetryPolicy | undefined;
+  retryPolicy: RetryPolicy;
   /**
    * Per-attempt ceiling, enforced alongside any caller `signal` rather than
    * instead of it. A cold local model can take a while to page in.
    */
-  timeoutMs: number | undefined;
+  timeoutMs: number;
   /**
    * Reads `Retry-After` off a failed response. Mirrors `ProviderAdapter`'s
    * member of the same name, so a provider that signals pacing its own way can
-   * override without touching the transport. Defaults to
-   * {@link extractRetryAfterMs}.
+   * override without touching the transport.
    */
-  extractRetryAfterMs: RetryAfterExtractor | undefined;
+  extractRetryAfterMs: RetryAfterExtractor;
   signal: AbortSignal | undefined;
 };
 
 type Attempt =
   | { ok: true; body: unknown }
   | { ok: false; error: InferenceError };
 
-const DEFAULT_TIMEOUT_MS = 30_000;
+export const DEFAULT_TIMEOUT_MS = 30_000;
 
 /**
  * `Retry-After` in seconds or as an HTTP date; undefined when absent.
@@ -88,7 +85,7 @@ const DEFAULT_TIMEOUT_MS = 30_000;
  * absolute instant, so it cannot be resolved against the harness scheduler's
  * virtual clock. Upstream's extractor has the same constraint.
  */
-const extractRetryAfterMs: RetryAfterExtractor = (headers) => {
+export const extractRetryAfterMs: RetryAfterExtractor = (headers) => {
   const header = headers.get("retry-after")?.trim();
   if (header === undefined || header === "") return undefined;
 
@@ -195,10 +192,8 @@ export async function runJSONRequest(
   deps: RequestDependencies,
   options: RunRequestOptions,
 ): Promise<unknown> {
-  const { signal } = options;
-  const policy = options.retryPolicy ?? createDefaultRetryPolicy();
-  const timeoutMs = options.timeoutMs ?? DEFAULT_TIMEOUT_MS;
-  const extractRetryAfter = options.extractRetryAfterMs ?? extractRetryAfterMs;
+  const { retryPolicy: policy, timeoutMs, signal } = options;
+  const extractRetryAfter = options.extractRetryAfterMs;
 
   // Time comes from the harness scheduler, not globals, so a virtual-clock
   // test scheduler drives the retry loop deterministically.