init

lib/ai/providers.ts

@@ -0,0 +1,58 @@
+import { createOpenAI } from '@ai-sdk/openai';
+import { getEncoding } from 'js-tiktoken';
+
+import { RecursiveCharacterTextSplitter } from './text-splitter';
+
+// Providers
+const openai = createOpenAI({
+  apiKey: import.meta.env.VITE_OPENAI_API_KEY!,
+  baseURL: import.meta.env.VITE_OPENAI_ENDPOINT || 'https://api.openai.com/v1',
+});
+
+const customModel = import.meta.env.VITE_OPENAI_MODEL || 'o3-mini';
+
+// Models
+
+export const o3MiniModel = openai(customModel, {
+  // reasoningEffort: customModel.startsWith('o') ? 'medium' : undefined,
+  structuredOutputs: true,
+});
+
+const MinChunkSize = 140;
+const encoder = getEncoding('o200k_base');
+
+// trim prompt to maximum context size
+export function trimPrompt(
+  prompt: string,
+  contextSize = Number(import.meta.env.VITE_CONTEXT_SIZE) || 128_000,
+) {
+  if (!prompt) {
+    return '';
+  }
+
+  const length = encoder.encode(prompt).length;
+  if (length <= contextSize) {
+    return prompt;
+  }
+
+  const overflowTokens = length - contextSize;
+  // on average it's 3 characters per token, so multiply by 3 to get a rough estimate of the number of characters
+  const chunkSize = prompt.length - overflowTokens * 3;
+  if (chunkSize < MinChunkSize) {
+    return prompt.slice(0, MinChunkSize);
+  }
+
+  const splitter = new RecursiveCharacterTextSplitter({
+    chunkSize,
+    chunkOverlap: 0,
+  });
+  const trimmedPrompt = splitter.splitText(prompt)[0] ?? '';
+
+  // last catch, there's a chance that the trimmed prompt is same length as the original prompt, due to how tokens are split & innerworkings of the splitter, handle this case by just doing a hard cut
+  if (trimmedPrompt.length === prompt.length) {
+    return trimPrompt(prompt.slice(0, chunkSize), contextSize);
+  }
+
+  // recursively trim until the prompt is within the context size
+  return trimPrompt(trimmedPrompt, contextSize);
+}