turns out kimi an others dont implement the new openai sdk

1 files changed, 86 insertions, 101 deletions

diff --git a/src/generic.ts b/src/generic.ts
index 3690dc6..e8dfa13 100644
--- a/src/generic.ts
+++ b/src/generic.ts
@@ -1,17 +1,16 @@
 import OpenAI from "openai";
 import { MAX_TOKENS, RESPONSE_LENGTH } from "./logic/constants";
-import type { AIModelAPI, InputToken } from "./types";
+import type { AIModelAPI, ChatMessage, InputToken } from "./types";
 import type { AsyncRes } from "sortug";
-import type {
-  ResponseCreateParamsBase,
-  ResponseCreateParamsNonStreaming,
-  ResponseCreateParamsStreaming,
-  ResponseInput,
-} from "openai/resources/responses/responses.mjs";
+import type { ChatCompletionContentPart } from "openai/resources";
+
+type OChoice = OpenAI.Chat.Completions.ChatCompletion.Choice;
+type Message = OpenAI.Chat.Completions.ChatCompletionUserMessageParam;
+type OMessage = OpenAI.Chat.Completions.ChatCompletionMessageParam;
 
 type Props = {
   baseURL: string;
-  apiKey: string | undefined;
+  apiKey: string;
   model?: string;
   maxTokens?: number;
   tokenizer?: (text: string) => number;
@@ -25,8 +24,6 @@ export default class OpenAIAPI implements AIModelAPI {
   model;
 
   constructor(props: Props) {
-    if (!props.apiKey) throw new Error("NO API KEY");
-    console.log({ props });
     this.apiKey = props.apiKey;
     this.baseURL = props.baseURL;
     this.api = new OpenAI({ baseURL: this.baseURL, apiKey: this.apiKey });
@@ -37,36 +34,43 @@ export default class OpenAIAPI implements AIModelAPI {
   public setModel(model: string) {
     this.model = model;
   }
-
-  public buildInput(tokens: InputToken[]): ResponseInput {
-    return [
-      {
-        role: "user",
-        content: tokens.map((t) =>
-          "text" in t
-            ? { type: "input_text", text: t.text }
-            : "img" in t
-              ? { type: "input_image", image_url: t.img, detail: "auto" }
-              : { type: "input_text", text: "oy vey" },
-        ),
-      },
-    ];
+  private mapMessages(input: ChatMessage[]): Message[] {
+    return input.map((m) => {
+      return { role: m.author as any, content: m.text, name: m.author };
+    });
+  }
+  private buildInput(tokens: InputToken[]): Message[] {
+    const content: ChatCompletionContentPart[] = tokens.map((t) => {
+      if ("text" in t) return { type: "text", text: t.text };
+      if ("img" in t) return { type: "image_url", image_url: { url: t.img } };
+      else return { type: "text", text: "oy vey" };
+    });
+    return [{ role: "user", content }];
   }
 
-  // OpenAI SDK has three kinds ReponseInputContent: text image and file
-  // images can be URLs or base64 dataurl thingies
-  //
   public async send(
-    inpt: string | InputToken[],
+    input: string | InputToken[],
     sys?: string,
   ): AsyncRes<string> {
-    const input = typeof inpt === "string" ? inpt : this.buildInput(inpt);
-    const params = sys ? { instructions: sys, input } : { input };
-    const res = await this.apiCall(params);
+    const messages: Message[] =
+      typeof input === "string"
+        ? [{ role: "user" as const, content: input }]
+        : this.buildInput(input);
+    // const messages = this.mapMessages(input);
+    const allMessages: OMessage[] = sys
+      ? [{ role: "system", content: sys }, ...messages]
+      : messages;
+    const truncated = this.truncateHistory(allMessages);
+    const res = await this.apiCall(truncated);
     if ("error" in res) return res;
     else {
       try {
-        return { ok: res.ok.output_text };
+        // TODO type this properly
+        const choices: OChoice[] = res.ok;
+        const resText = choices.reduce((acc, item) => {
+          return `${acc}\n${item.message.content || ""}`;
+        }, "");
+        return { ok: resText };
       } catch (e) {
         return { error: `${e}` };
       }
@@ -74,30 +78,46 @@ export default class OpenAIAPI implements AIModelAPI {
   }
 
   public async stream(
-    inpt: string | InputToken[],
+    input: string | InputToken[],
     handle: (c: string) => void,
     sys?: string,
   ) {
-    const input = typeof inpt === "string" ? inpt : this.buildInput(inpt);
-    const params = sys ? { instructions: sys, input } : { input };
-    await this.apiCallStream(params, handle);
+    const messages: Message[] =
+      typeof input === "string"
+        ? [{ role: "user" as const, content: input }]
+        : this.buildInput(input);
+    // const messages = this.mapMessages(input);
+    const allMessages: OMessage[] = sys
+      ? [{ role: "system", content: sys }, ...messages]
+      : messages;
+    const truncated = this.truncateHistory(allMessages);
+    await this.apiCallStream(truncated, handle);
+  }
+
+  private truncateHistory(messages: OMessage[]): OMessage[] {
+    const totalTokens = messages.reduce((total, message) => {
+      return total + this.tokenizer(message.content as string);
+    }, 0);
+    while (totalTokens > this.maxTokens && messages.length > 1) {
+      // Always keep the system message if it exists
+      const startIndex = messages[0].role === "system" ? 1 : 0;
+      messages.splice(startIndex, 1);
+    }
+    return messages;
   }
 
   // TODO custom temperature?
-  private async apiCall(
-    params: ResponseCreateParamsNonStreaming,
-  ): AsyncRes<OpenAI.Responses.Response> {
+  private async apiCall(messages: OMessage[]): AsyncRes<OChoice[]> {
+    // console.log({ messages }, "at the very end");
     try {
-      const res = await this.api.responses.create({
-        ...params,
+      const completion = await this.api.chat.completions.create({
         // temperature: 1.3,
-        model: params.model || this.model,
-        input: params.input,
-        max_output_tokens: params.max_output_tokens || RESPONSE_LENGTH,
-        stream: false,
+        model: this.model,
+        messages,
+        max_tokens: RESPONSE_LENGTH,
       });
-      // TODO damn there's a lot of stuff here
-      return { ok: res };
+      if (!completion) return { error: "null response from openai" };
+      return { ok: completion.choices };
     } catch (e) {
       console.log(e, "error in openai api");
       return { error: `${e}` };
@@ -105,65 +125,30 @@ export default class OpenAIAPI implements AIModelAPI {
   }
 
   private async apiCallStream(
-    params: ResponseCreateParamsBase,
-    handler: (c: string) => void,
-  ) {
-    // temperature: 1.3,
-    const pms: ResponseCreateParamsStreaming = {
-      ...params,
-      stream: true,
-      model: params.model || this.model,
-      input: params.input,
-      max_output_tokens: params.max_output_tokens || RESPONSE_LENGTH,
-    };
+    messages: OMessage[],
+    handle: (c: string) => void,
+  ): Promise<void> {
     try {
-      const stream = await this.api.responses.create(pms);
-      for await (const event of stream) {
-        console.log(event);
-        switch (event.type) {
-          // TODO deal with audio and whatever
-          case "response.output_text.delta":
-            handler(event.delta);
-            break;
-          case "response.completed":
-            break;
-          default:
-            break;
+      const stream = await this.api.chat.completions.create({
+        temperature: 1.3,
+        model: this.model,
+        messages,
+        max_tokens: RESPONSE_LENGTH,
+        stream: true,
+      });
+
+      for await (const chunk of stream) {
+        for (const choice of chunk.choices) {
+          console.log({ choice });
+          if (!choice.delta) continue;
+          const cont = choice.delta.content;
+          if (!cont) continue;
+          handle(cont);
         }
-        // if (event.type === "response.completed")
-        // wtf how do we use this
       }
     } catch (e) {
       console.log(e, "error in openai api");
-      return { error: `${e}` };
+      handle(`Error streaming OpenAI, ${e}`);
     }
   }
-
-  // private async apiCallStream(
-  //   messages: Message[],
-  //   handle: (c: string) => void,
-  // ): Promise<void> {
-  //   try {
-  //     const stream = await this.api.chat.completions.create({
-  //       temperature: 1.3,
-  //       model: this.model,
-  //       messages,
-  //       max_tokens: RESPONSE_LENGTH,
-  //       stream: true,
-  //     });
-
-  //     for await (const chunk of stream) {
-  //       for (const choice of chunk.choices) {
-  //         console.log({ choice });
-  //         if (!choice.delta) continue;
-  //         const cont = choice.delta.content;
-  //         if (!cont) continue;
-  //         handle(cont);
-  //       }
-  //     }
-  //   } catch (e) {
-  //     console.log(e, "error in openai api");
-  //     handle(`Error streaming OpenAI, ${e}`);
-  //   }
-  // }
 }