feat: relax chatFn types

examples/dexter/bin/election-news-chain.ts

@@ -16,6 +16,7 @@ async function main() {
   })
 
   const chain = createAIChain({
+    name: 'search_news',
     chatFn: chatModel.run.bind(chatModel),
     tools: [perigon.functions.pick('search_news_stories'), serper],
     params: {

packages/core/src/create-ai-chain.ts

@@ -80,6 +80,8 @@ export function createAIChain<Result extends types.AIChainResult = string>({
   const functionSet = new AIFunctionSet(tools)
   const schema = rawSchema ? asSchema(rawSchema, { strict }) : undefined
 
+  // TODO: support custom stopping criteria (like setting a flag in a tool call)
+
   const defaultParams: Partial<types.ChatParams> | undefined =
     schema && !functionSet.size
       ? {
@@ -190,10 +192,10 @@ export function createAIChain<Result extends types.AIChainResult = string>({
           throw new AbortError(
             'Function calls are not supported; expected tool call'
           )
+        } else if (Msg.isRefusal(message)) {
+          throw new AbortError(`Model refusal: ${message.refusal}`)
         } else if (Msg.isAssistant(message)) {
-          if (message.refusal) {
-            throw new AbortError(`Model refusal: ${message.refusal}`)
-          } else if (schema && schema.validate) {
+          if (schema && schema.validate) {
             const result = schema.validate(message.content)
 
             if (result.success) {
@@ -212,6 +214,8 @@ export function createAIChain<Result extends types.AIChainResult = string>({
           throw err
         }
 
+        console.warn(`Chain "${name}" error:`, err.message)
+
         messages.push(
           Msg.user(
             `There was an error validating the response. Please check the error message and try again.\nError:\n${getErrorMessage(err)}`

packages/core/src/message.ts

@@ -50,6 +50,15 @@ export interface Msg {
   name?: string
 }
 
+export interface LegacyMsg {
+  content: string | null
+  role: Msg.Role
+  function_call?: Msg.Call.Function
+  tool_calls?: Msg.Call.Tool[]
+  tool_call_id?: string
+  name?: string
+}
+
 /** Narrowed OpenAI Message types. */
 export namespace Msg {
   /** Possible roles for a message. */
@@ -102,8 +111,14 @@ export namespace Msg {
   export type Assistant = {
     role: 'assistant'
     name?: string
-    content?: string
-    refusal?: string
+    content: string
+  }
+
+  /** Message with refusal reason from the assistant. */
+  export type Refusal = {
+    role: 'assistant'
+    name?: string
+    refusal: string
   }
 
   /** Message with arguments to call a function. */
@@ -193,6 +208,27 @@ export namespace Msg {
     }
   }
 
+  /**
+   * Create an assistant refusal message. Cleans indentation and newlines by
+   * default.
+   */
+  export function refusal(
+    refusal: string,
+    opts?: {
+      /** Custom name for the message. */
+      name?: string
+      /** Whether to clean extra newlines and indentation. Defaults to true. */
+      cleanRefusal?: boolean
+    }
+  ): Msg.Refusal {
+    const { name, cleanRefusal = true } = opts ?? {}
+    return {
+      role: 'assistant',
+      refusal: cleanRefusal ? cleanStringForModel(refusal) : refusal,
+      ...(name ? { name } : {})
+    }
+  }
+
   /** Create a function call message with argumets. */
   export function funcCall(
     function_call: {
@@ -257,7 +293,7 @@ export namespace Msg {
     // @TODO
     response: any
     // response: ChatModel.EnrichedResponse
-  ): Msg.Assistant | Msg.FuncCall | Msg.ToolCall {
+  ): Msg.Assistant | Msg.Refusal | Msg.FuncCall | Msg.ToolCall {
     const msg = response.choices[0].message as Msg
     return narrowResponseMessage(msg)
   }
@@ -265,13 +301,15 @@ export namespace Msg {
   /** Narrow a message received from the API. It only responds with role=assistant */
   export function narrowResponseMessage(
     msg: Msg
-  ): Msg.Assistant | Msg.FuncCall | Msg.ToolCall {
+  ): Msg.Assistant | Msg.Refusal | Msg.FuncCall | Msg.ToolCall {
     if (msg.content === null && msg.tool_calls != null) {
       return Msg.toolCall(msg.tool_calls)
     } else if (msg.content === null && msg.function_call != null) {
       return Msg.funcCall(msg.function_call)
     } else if (msg.content !== null && msg.content !== undefined) {
       return Msg.assistant(msg.content)
+    } else if (msg.refusal != null) {
+      return Msg.refusal(msg.refusal)
     } else {
       // @TODO: probably don't want to error here
       console.log('Invalid message', msg)
@@ -291,6 +329,10 @@ export namespace Msg {
   export function isAssistant(message: Msg): message is Msg.Assistant {
     return message.role === 'assistant' && message.content !== null
   }
+  /** Check if a message is an assistant refusal message. */
+  export function isRefusal(message: Msg): message is Msg.Refusal {
+    return message.role === 'assistant' && message.refusal !== null
+  }
   /** Check if a message is a function call message with arguments. */
   export function isFuncCall(message: Msg): message is Msg.FuncCall {
     return message.role === 'assistant' && message.function_call != null
@@ -312,6 +354,7 @@ export namespace Msg {
   export function narrow(message: Msg.System): Msg.System
   export function narrow(message: Msg.User): Msg.User
   export function narrow(message: Msg.Assistant): Msg.Assistant
+  export function narrow(message: Msg.Assistant): Msg.Refusal
   export function narrow(message: Msg.FuncCall): Msg.FuncCall
   export function narrow(message: Msg.FuncResult): Msg.FuncResult
   export function narrow(message: Msg.ToolCall): Msg.ToolCall
@@ -322,6 +365,7 @@ export namespace Msg {
     | Msg.System
     | Msg.User
     | Msg.Assistant
+    | Msg.Refusal
     | Msg.FuncCall
     | Msg.FuncResult
     | Msg.ToolCall
@@ -335,6 +379,9 @@ export namespace Msg {
     if (isAssistant(message)) {
       return message
     }
+    if (isRefusal(message)) {
+      return message
+    }
     if (isFuncCall(message)) {
       return message
     }

packages/core/src/types.ts

@@ -3,7 +3,7 @@ import type { z } from 'zod'
 
 import type { AIFunctionSet } from './ai-function-set'
 import type { AIFunctionsProvider } from './fns'
-import type { Msg } from './message'
+import type { LegacyMsg, Msg } from './message'
 
 export type { Msg } from './message'
 export type { Schema } from './schema'
@@ -130,6 +130,10 @@ export interface ChatParams {
   user?: string
 }
 
+export type LegacyChatParams = Simplify<
+  Omit<ChatParams, 'messages'> & { messages: LegacyMsg[] }
+>
+
 export interface ResponseFormatJSONSchema {
   /**
    * The name of the response format. Must be a-z, A-Z, 0-9, or contain
@@ -158,10 +162,21 @@ export interface ResponseFormatJSONSchema {
   strict?: boolean
 }
 
+/**
+ * OpenAI has changed some of their types, so instead of trying to support all
+ * possible types, for these params, just relax them for now.
+ */
+export type RelaxedChatParams = Simplify<
+  Omit<ChatParams, 'messages' | 'response_format'> & {
+    messages: object[]
+    response_format?: { type: 'text' | 'json_object' | string }
+  }
+>
+
 /** An OpenAI-compatible chat completions API */
 export type ChatFn = (
-  params: Simplify<SetOptional<ChatParams, 'model'>>
-) => Promise<{ message: Msg }>
+  params: Simplify<SetOptional<RelaxedChatParams, 'model'>>
+) => Promise<{ message: Msg | LegacyMsg }>
 
 export type AIChainResult = string | Record<string, any>