@wmaurer/effect-ai-amazon-bedrock 0.1.0 → 0.3.0

package/src/AmazonBedrockLanguageModel.ts

@@ -23,7 +23,7 @@
  * deltas Converse returns are surfaced as document source parts.
  *
  * @stability unstable
- * @since 4.0.0
+ * @since 0.1.0
  */
 import * as AiError from "effect/ai/AiError"
 import { toCodecAnthropic } from "effect/ai/AnthropicStructuredOutput"
@@ -73,7 +73,7 @@ import * as InternalUtilities from "./internal/utilities.ts"
  *
  * @stability unstable
  * @category models
- * @since 4.0.0
+ * @since 0.1.0
  */
 export type Model = string & {}
 
@@ -92,7 +92,7 @@ export type Model = string & {}
  *
  * @stability unstable
  * @category models
- * @since 4.0.0
+ * @since 0.1.0
  */
 export type CachePoint = typeof CachePointBlock.Encoded
 
@@ -123,7 +123,7 @@ export type CachePoint = typeof CachePointBlock.Encoded
  *
  * @stability unstable
  * @category services
- * @since 4.0.0
+ * @since 0.1.0
  */
 export const ToolCachePoint = Context.Reference<CachePoint | undefined>(
   "@wmaurer/effect-ai-amazon-bedrock/AmazonBedrockLanguageModel/ToolCachePoint",
@@ -154,7 +154,7 @@ declare module "effect/ai/Prompt" {
    *
    * @stability unstable
    * @category models
-   * @since 4.0.0
+   * @since 0.1.0
    */
   export interface SystemMessageOptions extends CachePointOptions {}
 
@@ -163,7 +163,7 @@ declare module "effect/ai/Prompt" {
    *
    * @stability unstable
    * @category models
-   * @since 4.0.0
+   * @since 0.1.0
    */
   export interface UserMessageOptions extends CachePointOptions {}
 
@@ -172,7 +172,7 @@ declare module "effect/ai/Prompt" {
    *
    * @stability unstable
    * @category models
-   * @since 4.0.0
+   * @since 0.1.0
    */
   export interface AssistantMessageOptions extends CachePointOptions {}
 
@@ -181,7 +181,7 @@ declare module "effect/ai/Prompt" {
    *
    * @stability unstable
    * @category models
-   * @since 4.0.0
+   * @since 0.1.0
    */
   export interface ToolMessageOptions extends CachePointOptions {}
 
@@ -190,7 +190,7 @@ declare module "effect/ai/Prompt" {
    *
    * @stability unstable
    * @category models
-   * @since 4.0.0
+   * @since 0.1.0
    */
   export interface TextPartOptions extends CachePointOptions {}
 
@@ -199,7 +199,7 @@ declare module "effect/ai/Prompt" {
    *
    * @stability unstable
    * @category models
-   * @since 4.0.0
+   * @since 0.1.0
    */
   export interface ToolCallPartOptions extends CachePointOptions {}
 
@@ -208,7 +208,7 @@ declare module "effect/ai/Prompt" {
    *
    * @stability unstable
    * @category models
-   * @since 4.0.0
+   * @since 0.1.0
    */
   export interface ToolResultPartOptions extends CachePointOptions {}
 }
@@ -242,7 +242,7 @@ const pushCachePoint = (
  *
  * @stability unstable
  * @category models
- * @since 4.0.0
+ * @since 0.1.0
  */
 export type Citations = typeof CitationsConfig.Encoded
 
@@ -283,7 +283,7 @@ declare module "effect/ai/Prompt" {
    *
    * @stability unstable
    * @category models
-   * @since 4.0.0
+   * @since 0.1.0
    */
   export interface FilePartOptions extends FileOptions {}
 }
@@ -299,7 +299,7 @@ declare module "effect/ai/Response" {
    *
    * @stability unstable
    * @category models
-   * @since 4.0.0
+   * @since 0.1.0
    */
   export interface DocumentSourcePartMetadata extends ProviderMetadata {
     readonly amazonBedrock?: {
@@ -403,7 +403,7 @@ const citationSource: (
  *
  * @stability unstable
  * @category models
- * @since 4.0.0
+ * @since 0.1.0
  */
 export type ReasoningInfo = {
   readonly type: "reasoningText"
@@ -433,7 +433,7 @@ declare module "effect/ai/Prompt" {
    *
    * @stability unstable
    * @category models
-   * @since 4.0.0
+   * @since 0.1.0
    */
   export interface ReasoningPartOptions extends ProviderOptions {
     readonly amazonBedrock?: {
@@ -448,7 +448,7 @@ declare module "effect/ai/Response" {
    *
    * @stability unstable
    * @category models
-   * @since 4.0.0
+   * @since 0.1.0
    */
   export interface ReasoningPartMetadata extends ProviderMetadata {
     readonly amazonBedrock?: {
@@ -467,7 +467,7 @@ declare module "effect/ai/Response" {
    *
    * @stability unstable
    * @category models
-   * @since 4.0.0
+   * @since 0.1.0
    */
   export interface ReasoningDeltaPartMetadata extends ProviderMetadata {
     readonly amazonBedrock?: {
@@ -525,7 +525,7 @@ const reasoningTextInfo = (signature: string | undefined): ReasoningInfo => ({
  *
  * @stability unstable
  * @category configuration
- * @since 4.0.0
+ * @since 0.1.0
  */
 export class Config extends Context.Service<
   Config,
@@ -554,11 +554,15 @@ export class Config extends Context.Service<
        *
        * **Details**
        *
-       * Overrides automatic capability detection based on the model
-       * identifier. When `false`, `generateObject` asks for the object through
-       * a tool instead, which is forced on models that accept a forced
-       * `toolChoice` and left unforced, with an instruction to call it, on
-       * models that do not. Never sent to Bedrock.
+       * Overrides the default for the provider that the model identifier
+       * names. By default only Claude models, and ARNs that do not name their
+       * model, get native structured output. In probes on 2026-10-17 in
+       * eu-west-1 and us-east-1, Nova, Mistral Large 2402, Pixtral Large,
+       * Llama, and Palmyra answered `outputConfig` with a 400. When `false`,
+       * `generateObject` asks for the object through a tool instead, which is
+       * forced on models that accept a forced `toolChoice` and left unforced,
+       * with an instruction to call it, on models that do not. Never sent to
+       * Bedrock.
        *
        * When the prompt holds tool calls or tool results and the request
        * offers no tools, `generateObject` uses the tool even when this is
@@ -571,12 +575,19 @@ export class Config extends Context.Service<
        *
        * **Details**
        *
-       * Overrides automatic capability detection based on the model
-       * identifier. It decides whether the tool that `generateObject` uses
-       * without native structured output is forced, so with
-       * `structuredOutputs: false` it selects between the forced and the
-       * unforced tool, for example for an application inference profile ARN,
-       * which does not name its model. Never sent to Bedrock.
+       * Overrides the default for the provider that the model identifier
+       * names. It decides whether the tool that `generateObject` uses
+       * without native structured output is forced with `toolChoice`
+       * `{ any: {} }`, so with `structuredOutputs: false` it selects between
+       * the forced and the unforced tool, for example for an application
+       * inference profile ARN, which does not name its model. Never sent to
+       * Bedrock.
+       *
+       * By default Meta and Writer models, and Claude Opus 5.5, Sonnet 5.5,
+       * Fable 5.1, and Mythos 5.1, get the unforced tool. In probes on
+       * 2026-10-17 in us-east-1, Llama 3.1, 3.3, and 4 and Palmyra X4 and X5
+       * answered `any` with a 400, and in eu-west-1 Opus 5.5 and Sonnet 5.5
+       * rejected both `any` and a named `tool`.
        *
        * It steers only that tool. A `toolChoice` passed with the caller's own
        * tools, such as `"required"` or a named tool, goes to Bedrock
@@ -598,7 +609,7 @@ export class Config extends Context.Service<
  *
  * @stability unstable
  * @category constructors
- * @since 4.0.0
+ * @since 0.1.0
  */
 export const model = (
   model: (string & {}) | Model,
@@ -612,7 +623,7 @@ export const model = (
  *
  * @stability unstable
  * @category constructors
- * @since 4.0.0
+ * @since 0.1.0
  */
 export const make = Effect.fnUntraced(function*({ config: providerConfig, model }: {
   readonly model: (string & {}) | Model
@@ -674,9 +685,9 @@ export const make = Effect.fnUntraced(function*({ config: providerConfig, model
             }
           }
         } else {
-          // The object is asked for through a tool, which replaces the
-          // caller's tools. A model that rejects a forced `toolChoice` gets
-          // the tool unforced, and the request says to call it.
+          // The object tool is the only tool, so `any` forces it, and
+          // Mistral Large 2402 and Pixtral Large accept `any` but reject a
+          // named `tool`.
           const forced = forcedToolChoice ?? capabilities.supportsForcedToolChoice
           nameMapper = new Tool.NameMapper([])
           toolConfig = {
@@ -687,7 +698,7 @@ export const make = Effect.fnUntraced(function*({ config: providerConfig, model
                 inputSchema: { json: json as Record<string, unknown> }
               }
             }],
-            toolChoice: forced ? { tool: { name: responseFormat.objectName } } : { auto: {} }
+            toolChoice: forced ? { any: {} } : { auto: {} }
           }
           if (!forced) {
             system = [...(system ?? []), {
@@ -742,7 +753,7 @@ export const make = Effect.fnUntraced(function*({ config: providerConfig, model
  *
  * @stability unstable
  * @category layers
- * @since 4.0.0
+ * @since 0.1.0
  */
 export const layer = (options: {
   readonly model: (string & {}) | Model
@@ -755,7 +766,7 @@ export const layer = (options: {
  *
  * @stability unstable
  * @category configuration
- * @since 4.0.0
+ * @since 0.1.0
  */
 export const withConfigOverride: {
   (overrides: typeof Config.Service): <A, E, R>(self: Effect.Effect<A, E, R>) => Effect.Effect<A, E, Exclude<R, Config>>
@@ -1124,16 +1135,36 @@ interface ModelCapabilities {
 }
 
 /**
- * Returns what a model accepts for structured output, optimistically assuming
- * modern capabilities for unknown IDs. Matched by substring, so foundation
- * model ids and inference profile ids (`eu.anthropic.…`) both resolve.
+ * Returns what a model accepts for structured output, looked up by the
+ * provider that the model ID names: `anthropic` in
+ * `eu.anthropic.claude-opus-4-8`, `amazon` in `amazon.nova-pro-v1:0`. A
+ * foundation model or inference profile ARN names its model at the end. Other
+ * ARNs, such as an application inference profile or a provisioned model, do
+ * not. They get the Claude rules, which give them the native structured output
+ * and forced tool of 0.2.0, on the unmeasured assumption that most of them
+ * wrap a Claude model.
+ *
+ * Probes on 2026-10-17, one run per model and call unless noted:
  *
+ * - Only Claude gets native structured output. Nova, Mistral Large 2402 and
+ *   Pixtral Large in eu-west-1, and Llama, Palmyra, DeepSeek R1, Mistral 7B
+ *   and Mixtral in us-east-1, answer `outputConfig` with a 400. gpt-oss-120b
+ *   accepts it but wrote text before the JSON in 3 of 4 runs.
+ * - The forced object tool, sent as `{ any: {} }`, returned the object on 15
+ *   of 16 models in eu-west-1, Claude Haiku 4.5 and Opus 4.8 included, and
+ *   on all 17 models in us-east-1 that accept a forced tool. Magistral Small
+ *   fails every object path.
+ * - Llama 3.1, 3.3 and 4 and Palmyra X4 and X5 (`meta`, `writer`) answer
+ *   `any` with a 400 in us-east-1, and so does each of them that the probe
+ *   sent a named `tool`, all but Llama 3.1 8B. They get the unforced tool. It
+ *   returned the object on Llama 3.1 8B, Llama 4 Scout and Maverick and
+ *   Palmyra X5, and a wrong object on Llama 3.1 70B and 3.3 70B.
  * - Claude models before Opus 4.1 / Sonnet 4.5 predate structured output, as
  *   in the Anthropic provider. This is not checked on Bedrock: a probe of 17
- *   Claude inference profiles in eu-west-1 on 2026-10-17 found Claude 3 Haiku,
- *   Claude 3 Sonnet and Sonnet 4 at end of life there, and did not try Claude
- *   3 Opus, 3.5 Haiku, 3.5 Sonnet, 3.7 Sonnet, Opus 4 or Opus 4.1.
- * - Every Claude profile that served the probe's requests (12 of 17)
+ *   Claude inference profiles in eu-west-1 found Claude 3 Haiku, Claude 3
+ *   Sonnet and Sonnet 4 at end of life there, and did not try Claude 3 Opus,
+ *   3.5 Haiku, 3.5 Sonnet, 3.7 Sonnet, Opus 4 or Opus 4.1.
+ * - Every Claude profile that served that probe's requests (12 of 17)
  *   accepted `outputConfig.textFormat`: Haiku 4.5 and 5.5, Opus 4.5 to 4.8, 5
  *   and 5.5, and Sonnet 4.5, 4.6, 5 and 5.5. Fable 5 and 5.1 rejected every
  *   request, with or without `textFormat`, with a data-retention
@@ -1146,23 +1177,51 @@ interface ModelCapabilities {
  * @see https://platform.claude.com/docs/en/build-with-claude/structured-outputs
  */
 const getModelCapabilities = (modelId: string): ModelCapabilities => {
-  const supportsForcedToolChoice = !(
-    modelId.includes("claude-opus-5-5") ||
-    modelId.includes("claude-sonnet-5-5") ||
-    modelId.includes("claude-fable-5-1") ||
-    modelId.includes("claude-mythos-5-1")
-  )
-  const supportsStructuredOutput = !(
-    modelId.includes("claude-sonnet-4-0") ||
-    modelId.includes("claude-sonnet-4-20250514") ||
-    modelId.includes("claude-3-7-sonnet") ||
-    modelId.includes("claude-opus-4-0") ||
-    modelId.includes("claude-opus-4-20250514") ||
-    modelId.includes("claude-3-")
-  )
-  return { supportsStructuredOutput, supportsForcedToolChoice }
+  const named = modelId.startsWith("arn:") ? namingArn.exec(modelId)?.[1] : modelId
+  if (Predicate.isUndefined(named)) {
+    return claudeCapabilities(modelId)
+  }
+  // `provider.model`, or `geo.provider.model` for an inference profile.
+  const [first = "", second = ""] = named.split(".")
+  const byProvider = capabilitiesByProvider.get(first) ?? capabilitiesByProvider.get(second)
+  return Predicate.isUndefined(byProvider) ? otherProvider : byProvider(named)
 }
 
+// Of the ARNs Converse accepts, only these two end in the ID of the model
+// they invoke.
+const namingArn = /^arn:[^:]+:bedrock:[^:]*:[^:]*:(?:foundation-model|inference-profile)\/(.+)$/
+
+const otherProvider: ModelCapabilities = { supportsStructuredOutput: false, supportsForcedToolChoice: true }
+
+const unforcedTool: ModelCapabilities = { supportsStructuredOutput: false, supportsForcedToolChoice: false }
+
+const claudeWithoutStructuredOutput = [
+  "claude-sonnet-4-0",
+  "claude-sonnet-4-20250514",
+  "claude-3-7-sonnet",
+  "claude-opus-4-0",
+  "claude-opus-4-20250514",
+  "claude-3-"
+]
+
+const claudeWithoutForcedToolChoice = [
+  "claude-opus-5-5",
+  "claude-sonnet-5-5",
+  "claude-fable-5-1",
+  "claude-mythos-5-1"
+]
+
+const claudeCapabilities = (modelId: string): ModelCapabilities => ({
+  supportsStructuredOutput: !claudeWithoutStructuredOutput.some((id) => modelId.includes(id)),
+  supportsForcedToolChoice: !claudeWithoutForcedToolChoice.some((id) => modelId.includes(id))
+})
+
+const capabilitiesByProvider = new Map<string, (modelId: string) => ModelCapabilities>([
+  ["anthropic", claudeCapabilities],
+  ["meta", () => unforcedTool],
+  ["writer", () => unforcedTool]
+])
+
 const tryJsonSchema = (schema: Schema.Top, method: string) =>
   Effect.try({
     try: () => Tool.getJsonSchemaFromSchema(schema, { transformer: toCodecAnthropic }),
@@ -1591,6 +1650,19 @@ const makeStreamResponse: (
             const block = toolBlocks.get(index)
             if (Predicate.isNotUndefined(block)) {
               toolBlocks.delete(index)
+              const params = yield* Effect.try({
+                // A tool taking no arguments streams no deltas at all.
+                try: () => Tool.unsafeSecureJsonParse(block.params.length === 0 ? "{}" : block.params),
+                catch: () =>
+                  AiError.make({
+                    module: "AmazonBedrockLanguageModel",
+                    method: "makeStreamResponse",
+                    reason: new AiError.InvalidOutputError({
+                      description: `Bedrock streamed input for the \`${block.name}\` tool that is not valid JSON: ` +
+                        truncate(block.params, 500)
+                    })
+                  })
+              })
               parts.push({
                 type: "tool-params-end",
                 id: block.id
@@ -1599,8 +1671,7 @@ const makeStreamResponse: (
                 type: "tool-call",
                 id: block.id,
                 name: block.name,
-                // A tool taking no arguments streams no deltas at all.
-                params: Tool.unsafeSecureJsonParse(block.params.length === 0 ? "{}" : block.params),
+                params,
                 providerExecuted: false
               })
             }
@@ -1744,3 +1815,6 @@ const trimIfLast = (
   isLastPart: boolean,
   text: string
 ) => isLastGroup && isLastMessage && isLastPart ? text.trim() : text
+
+const truncate = (text: string, maxLength: number): string =>
+  text.length > maxLength ? `${text.slice(0, maxLength)}...` : text