@juspay/neurolink 12.49.13 → 12.49.15

package/dist/utils/tokenLimits.js

@@ -2,7 +2,8 @@
  * Provider-specific token limit utilities
  * Provides safe maxTokens values based on provider and model capabilities
  */
-import { getRuntimeOutputCeiling, stripBedrockGeoPrefix, } from "../constants/contextWindows.js";
+import { getRuntimeOutputCeiling } from "../constants/contextWindows.js";
+import { bedrockModelIdForLookup, stripBedrockGeoPrefix, } from "../constants/bedrockIdentifiers.js";
 import { PROVIDER_MAX_TOKENS } from "../core/constants.js";
 import { logger } from "./logger.js";
 import { hasRestrictedOutputLimit, RESTRICTED_OUTPUT_TOKEN_LIMIT, } from "./modelDetection.js";
@@ -35,6 +36,11 @@ export function getSafeMaxTokens(provider, model, requestedMaxTokens) {
         // Otherwise, use the requested value (it's within limits, including 0)
         return requestedMaxTokens;
     }
+    // The descriptor registers aws as Bedrock. Options still carry the caller
+    // spelling when this normalizer runs, before provider instantiation.
+    const canonicalProvider = /^(?:bedrock|aws)$/i.test(provider)
+        ? "bedrock"
+        : provider;
     // Runtime-discovered output ceiling (e.g. LiteLLM /model/info
     // max_output_tokens): the serving infrastructure's own number for the
     // deployed model, authoritative over the static per-provider table. This
@@ -44,7 +50,10 @@ export function getSafeMaxTokens(provider, model, requestedMaxTokens) {
     // usable input on total-context backends. Checked AFTER the
     // restricted-model branch so hard caps proven at the origin API (Gemini 3
     // / image models) still win over a proxy that over-advertises them.
-    const runtimeCeiling = getRuntimeOutputCeiling(provider, model);
+    const runtimeCeiling = getRuntimeOutputCeiling(provider, model) ??
+        (canonicalProvider === "bedrock"
+            ? getRuntimeOutputCeiling(canonicalProvider, model)
+            : undefined);
     if (runtimeCeiling !== undefined) {
         if (requestedMaxTokens === undefined || requestedMaxTokens === null) {
             return runtimeCeiling;
@@ -57,7 +66,7 @@ export function getSafeMaxTokens(provider, model, requestedMaxTokens) {
         return requestedMaxTokens;
     }
     // Get provider-specific limits
-    const providerLimits = PROVIDER_MAX_TOKENS[provider];
+    const providerLimits = PROVIDER_MAX_TOKENS[canonicalProvider];
     if (!providerLimits) {
         logger.warn(`Unknown provider ${provider}, no token limits enforced`);
         // Explicit undefined/null check so a caller-supplied 0 is preserved
@@ -71,7 +80,12 @@ export function getSafeMaxTokens(provider, model, requestedMaxTokens) {
     // profile id (us.amazon.nova-pro-v1:0) is the only one a caller can invoke.
     // The bare id is tried only after the id as given misses, so an explicit row
     // for the prefixed id still wins.
-    const bareModel = provider === "bedrock" && model ? stripBedrockGeoPrefix(model) : undefined;
+    const lookupModel = canonicalProvider === "bedrock" && model
+        ? bedrockModelIdForLookup(model)
+        : undefined;
+    const bareModel = lookupModel
+        ? stripBedrockGeoPrefix(lookupModel)
+        : undefined;
     // Get model-specific limit or provider default
     let maxLimit;
     if (model &&
@@ -79,6 +93,11 @@ export function getSafeMaxTokens(provider, model, requestedMaxTokens) {
         providerLimits[model]) {
         maxLimit = providerLimits[model];
     }
+    else if (lookupModel !== undefined &&
+        typeof providerLimits === "object" &&
+        providerLimits[lookupModel]) {
+        maxLimit = providerLimits[lookupModel];
+    }
     else if (bareModel !== undefined &&
         typeof providerLimits === "object" &&
         providerLimits[bareModel]) {