😼

examples/functions.ts

@@ -9,7 +9,7 @@ async function main() {
   const agentic = new Agentic({ openai })
 
   const example = await agentic
-    .gpt4('What is 5 * 50?')
+    .gpt3('What is 5 * 50?')
     .tools([new CalculatorTool({ agentic })])
     .output(z.object({ answer: z.number() }))
     .call()

src/agentic.ts

@@ -1,6 +1,6 @@
 import * as types from '@/types'
 import { DEFAULT_OPENAI_MODEL } from '@/constants'
-import { OpenAIChatModel } from '@/llms/openai'
+import { OpenAIChatCompletion } from '@/llms/openai'
 
 import {
   HumanFeedbackMechanism,
@@ -77,7 +77,7 @@ export class Agentic {
     return this._idGeneratorFn
   }
 
-  llm(
+  openaiChatCompletion(
     promptOrChatCompletionParams:
       | string
       | Partial<types.openai.ChatCompletionParams> // TODO: make more strict
@@ -101,13 +101,16 @@ export class Agentic {
       }
     }
 
-    return new OpenAIChatModel({
+    return new OpenAIChatCompletion({
       agentic: this,
       ...(this._openaiModelDefaults as any), // TODO
       ...options
     })
   }
 
+  /**
+   * Shortcut for creating an OpenAI chat completion call with the `gpt-3.5-turbo` model.
+   */
   gpt3(
     promptOrChatCompletionParams:
       | string
@@ -132,7 +135,7 @@ export class Agentic {
       }
     }
 
-    return new OpenAIChatModel({
+    return new OpenAIChatCompletion({
       agentic: this,
       ...(this._openaiModelDefaults as any), // TODO
       model: 'gpt-3.5-turbo',
@@ -140,6 +143,9 @@ export class Agentic {
     })
   }
 
+  /**
+   * Shortcut for creating an OpenAI chat completion call with the `gpt-4` model.
+   */
   gpt4(
     promptOrChatCompletionParams:
       | string
@@ -164,7 +170,7 @@ export class Agentic {
       }
     }
 
-    return new OpenAIChatModel({
+    return new OpenAIChatCompletion({
       agentic: this,
       ...(this._openaiModelDefaults as any), // TODO
       model: 'gpt-4',

src/llms/anthropic.ts

@@ -4,14 +4,14 @@ import { type SetOptional } from 'type-fest'
 import * as types from '@/types'
 import { DEFAULT_ANTHROPIC_MODEL } from '@/constants'
 
-import { BaseChatModel } from './chat'
+import { BaseChatCompletion } from './chat'
 
 const defaultStopSequences = [anthropic.HUMAN_PROMPT]
 
-export class AnthropicChatModel<
+export class AnthropicChatCompletion<
   TInput extends void | types.JsonObject = any,
   TOutput extends types.JsonValue = string
-> extends BaseChatModel<
+> extends BaseChatCompletion<
   TInput,
   TOutput,
   SetOptional<
@@ -42,7 +42,7 @@ export class AnthropicChatModel<
       this._client = this._agentic.anthropic
     } else {
       throw new Error(
-        'AnthropicChatModel requires an Anthropic client to be configured on the Agentic runtime'
+        'AnthropicChatCompletion requires an Anthropic client to be configured on the Agentic runtime'
       )
     }
   }
@@ -91,8 +91,8 @@ export class AnthropicChatModel<
     }
   }
 
-  public override clone(): AnthropicChatModel<TInput, TOutput> {
+  public override clone(): AnthropicChatCompletion<TInput, TOutput> {
-    return new AnthropicChatModel<TInput, TOutput>({
+    return new AnthropicChatCompletion<TInput, TOutput>({
       agentic: this._agentic,
       timeoutMs: this._timeoutMs,
       retryConfig: this._retryConfig,

src/llms/chat.ts

@@ -19,7 +19,7 @@ import {
   getNumTokensForChatMessages
 } from './llm-utils'
 
-export abstract class BaseChatModel<
+export abstract class BaseChatCompletion<
   TInput extends void | types.JsonObject = void,
   TOutput extends types.JsonValue = string,
   TModelParams extends Record<string, any> = Record<string, any>,
@@ -43,8 +43,8 @@ export abstract class BaseChatModel<
   // TODO: use polymorphic `this` type to return correct BaseLLM subclass type
   input<U extends void | types.JsonObject>(
     inputSchema: ZodType<U>
-  ): BaseChatModel<U, TOutput, TModelParams> {
+  ): BaseChatCompletion<U, TOutput, TModelParams> {
-    const refinedInstance = this as unknown as BaseChatModel<
+    const refinedInstance = this as unknown as BaseChatCompletion<
       U,
       TOutput,
       TModelParams
@@ -56,8 +56,8 @@ export abstract class BaseChatModel<
   // TODO: use polymorphic `this` type to return correct BaseLLM subclass type
   output<U extends types.JsonValue>(
     outputSchema: ZodType<U>
-  ): BaseChatModel<TInput, U, TModelParams> {
+  ): BaseChatCompletion<TInput, U, TModelParams> {
-    const refinedInstance = this as unknown as BaseChatModel<
+    const refinedInstance = this as unknown as BaseChatCompletion<
       TInput,
       U,
       TModelParams
@@ -66,11 +66,9 @@ export abstract class BaseChatModel<
     return refinedInstance
   }
 
-  tools(tools: BaseTask<any, any>[]): this {
+  public tools(tools: BaseTask<any, any>[]): this {
     if (!this.supportsTools) {
-      throw new Error(
+      throw new Error(`This LLM "${this.nameForHuman}" does not support tools`)
-        `This Chat model "${this.nameForHuman}" does not support tools`
-      )
     }
 
     this._tools = tools
@@ -179,20 +177,30 @@ export abstract class BaseChatModel<
       }
     }
 
+    let output: any
+
+    do {
+      console.log('<<< completion', { messages, functions })
       const completion = await this._createChatCompletion(messages, functions)
+      console.log('>>> completion', completion.message)
       ctx.metadata.completion = completion
 
       if (completion.message.function_call) {
         const functionCall = completion.message.function_call
 
-      if (isUsingTools) {
+        if (!isUsingTools) {
+          // TODO: not sure what we should do in this case...
+          output = functionCall
+          break
+        }
+
         const tool = this._tools!.find(
           (tool) => tool.nameForModel === functionCall.name
         )
 
         if (!tool) {
           throw new errors.OutputValidationError(
-            `Unrecognized function call "${functionCall.name}"`
+            `Function not found "${functionCall.name}"`
           )
         }
 
@@ -201,7 +209,9 @@ export abstract class BaseChatModel<
           functionArguments = JSON.parse(jsonrepair(functionCall.arguments))
         } catch (err: any) {
           if (err instanceof JSONRepairError) {
-            throw new errors.OutputValidationError(err.message, { cause: err })
+            throw new errors.OutputValidationError(err.message, {
+              cause: err
+            })
           } else if (err instanceof SyntaxError) {
             throw new errors.OutputValidationError(
               `Invalid JSON object: ${err.message}`,
@@ -212,11 +222,22 @@ export abstract class BaseChatModel<
           }
         }
 
+        console.log('>>> sub-task', {
+          task: functionCall.name,
+          input: functionArguments
+        })
+
         // TODO: handle sub-task errors gracefully
         const toolCallResponse = await tool.callWithMetadata(functionArguments)
 
+        console.log('<<< sub-task', {
+          task: functionCall.name,
+          input: functionArguments,
+          output: toolCallResponse.result
+        })
+
         // TODO: handle result as string or JSON
-        // TODO:
+        // TODO: better support for JSON spacing
         const taskCallContent = JSON.stringify(
           toolCallResponse.result,
           null,
@@ -231,12 +252,11 @@ export abstract class BaseChatModel<
           content: taskCallContent
         })
 
-        // TODO: re-invoke completion with new messages
+        continue
-        throw new Error('TODO')
-      }
       }
 
-    let output: any = completion.message.content
+      output = completion.message.content
+    } while (output === undefined)
 
     console.log('===')
     console.log(output)

src/llms/openai.ts

@@ -2,8 +2,9 @@ import type { SetOptional } from 'type-fest'
 
 import * as types from '@/types'
 import { DEFAULT_OPENAI_MODEL } from '@/constants'
+import { BaseTask } from '@/task'
 
-import { BaseChatModel } from './chat'
+import { BaseChatCompletion } from './chat'
 
 const openaiModelsSupportingFunctions = new Set([
   'gpt-4-0613',
@@ -12,10 +13,10 @@ const openaiModelsSupportingFunctions = new Set([
   'gpt-3.5-turbo-16k'
 ])
 
-export class OpenAIChatModel<
+export class OpenAIChatCompletion<
   TInput extends void | types.JsonObject = any,
   TOutput extends types.JsonValue = string
-> extends BaseChatModel<
+> extends BaseChatCompletion<
   TInput,
   TOutput,
   SetOptional<Omit<types.openai.ChatCompletionParams, 'messages'>, 'model'>,
@@ -41,7 +42,7 @@ export class OpenAIChatModel<
       this._client = this._agentic.openai
     } else {
       throw new Error(
-        'OpenAIChatModel requires an OpenAI client to be configured on the Agentic runtime'
+        'OpenAIChatCompletion requires an OpenAI client to be configured on the Agentic runtime'
       )
     }
 
@@ -65,7 +66,27 @@ export class OpenAIChatModel<
   }
 
   public override get nameForHuman(): string {
-    return `OpenAIChatModel ${this._model}`
+    return `OpenAIChatCompletion ${this._model}`
+  }
+
+  public override tools(tools: BaseTask<any, any>[]): this {
+    if (!this.supportsTools) {
+      switch (this._model) {
+        case 'gpt-4':
+          this._model = 'gpt-4-0613'
+          break
+
+        case 'gpt-4-32k':
+          this._model = 'gpt-4-32k-0613'
+          break
+
+        case 'gpt-3.5-turbo':
+          this._model = 'gpt-3.5-turbo-0613'
+          break
+      }
+    }
+
+    return super.tools(tools)
   }
 
   public override get supportsTools(): boolean {
@@ -86,8 +107,8 @@ export class OpenAIChatModel<
     })
   }
 
-  public override clone(): OpenAIChatModel<TInput, TOutput> {
+  public override clone(): OpenAIChatCompletion<TInput, TOutput> {
-    return new OpenAIChatModel<TInput, TOutput>({
+    return new OpenAIChatCompletion<TInput, TOutput>({
       agentic: this._agentic,
       timeoutMs: this._timeoutMs,
       retryConfig: this._retryConfig,

test/llms/anthropic.test.ts

@@ -1,15 +1,15 @@
 import test from 'ava'
 import { expectTypeOf } from 'expect-type'
 
-import { AnthropicChatModel } from '@/llms/anthropic'
+import { AnthropicChatCompletion } from '@/llms/anthropic'
 
 import { createTestAgenticRuntime } from '../_utils'
 
-test('AnthropicChatModel ⇒ string output', async (t) => {
+test('AnthropicChatCompletion ⇒ string output', async (t) => {
   t.timeout(2 * 60 * 1000)
   const agentic = createTestAgenticRuntime()
 
-  const builder = new AnthropicChatModel({
+  const builder = new AnthropicChatCompletion({
     agentic,
     modelParams: {
       temperature: 0,

test/llms/openai.test.ts

@@ -4,16 +4,16 @@ import sinon from 'sinon'
 import { z } from 'zod'
 
 import { OutputValidationError, TemplateValidationError } from '@/errors'
-import { BaseChatModel, OpenAIChatModel } from '@/llms'
+import { BaseChatCompletion, OpenAIChatCompletion } from '@/llms'
 
 import { createTestAgenticRuntime } from '../_utils'
 
-test('OpenAIChatModel ⇒ types', async (t) => {
+test('OpenAIChatCompletion ⇒ types', async (t) => {
   const agentic = createTestAgenticRuntime()
   const b = agentic.gpt4('test')
   t.pass()
 
-  expectTypeOf(b).toMatchTypeOf<OpenAIChatModel<any, string>>()
+  expectTypeOf(b).toMatchTypeOf<OpenAIChatCompletion<any, string>>()
 
   expectTypeOf(
     b.input(
@@ -22,7 +22,7 @@ test('OpenAIChatModel ⇒ types', async (t) => {
       })
     )
   ).toMatchTypeOf<
-    BaseChatModel<
+    BaseChatCompletion<
       {
         foo: string
       },
@@ -37,7 +37,7 @@ test('OpenAIChatModel ⇒ types', async (t) => {
       })
     )
   ).toMatchTypeOf<
-    BaseChatModel<
+    BaseChatCompletion<
       any,
       {
         bar?: string
@@ -46,11 +46,11 @@ test('OpenAIChatModel ⇒ types', async (t) => {
   >()
 })
 
-test('OpenAIChatModel ⇒ string output', async (t) => {
+test('OpenAIChatCompletion ⇒ string output', async (t) => {
   t.timeout(2 * 60 * 1000)
   const agentic = createTestAgenticRuntime()
 
-  const builder = new OpenAIChatModel({
+  const builder = new OpenAIChatCompletion({
     agentic,
     modelParams: {
       temperature: 0,
@@ -80,11 +80,11 @@ test('OpenAIChatModel ⇒ string output', async (t) => {
   expectTypeOf(result2).toMatchTypeOf<string>()
 })
 
-test('OpenAIChatModel ⇒ json output', async (t) => {
+test('OpenAIChatCompletion ⇒ json output', async (t) => {
   t.timeout(2 * 60 * 1000)
   const agentic = createTestAgenticRuntime()
 
-  const builder = new OpenAIChatModel({
+  const builder = new OpenAIChatCompletion({
     agentic,
     modelParams: {
       temperature: 0.5
@@ -106,11 +106,11 @@ test('OpenAIChatModel ⇒ json output', async (t) => {
   expectTypeOf(result).toMatchTypeOf<{ foo: string; bar: number }>()
 })
 
-test('OpenAIChatModel ⇒ boolean output', async (t) => {
+test('OpenAIChatCompletion ⇒ boolean output', async (t) => {
   t.timeout(2 * 60 * 1000)
   const agentic = createTestAgenticRuntime()
 
-  const builder = new OpenAIChatModel({
+  const builder = new OpenAIChatCompletion({
     agentic,
     modelParams: {
       temperature: 0,
@@ -130,11 +130,11 @@ test('OpenAIChatModel ⇒ boolean output', async (t) => {
   expectTypeOf(result).toMatchTypeOf<boolean>()
 })
 
-test('OpenAIChatModel ⇒ retry logic', async (t) => {
+test('OpenAIChatCompletion ⇒ retry logic', async (t) => {
   t.timeout(2 * 60 * 1000)
   const agentic = createTestAgenticRuntime()
 
-  const builder = new OpenAIChatModel({
+  const builder = new OpenAIChatCompletion({
     agentic,
     modelParams: {
       temperature: 0,
@@ -167,7 +167,7 @@ test('OpenAIChatModel ⇒ retry logic', async (t) => {
   t.is(fakeCall.callCount, 3)
 })
 
-test('OpenAIChatModel ⇒ template variables', async (t) => {
+test('OpenAIChatCompletion ⇒ template variables', async (t) => {
   t.timeout(2 * 60 * 1000)
   const agentic = createTestAgenticRuntime()
 
@@ -203,7 +203,7 @@ test('OpenAIChatModel ⇒ template variables', async (t) => {
   }
 })
 
-test('OpenAIChatModel ⇒ missing template variable', async (t) => {
+test('OpenAIChatCompletion ⇒ missing template variable', async (t) => {
   t.timeout(2 * 60 * 1000)
   const agentic = createTestAgenticRuntime()