diff options
Diffstat (limited to 'app/src/lib/llm-service.ts')
| -rw-r--r-- | app/src/lib/llm-service.ts | 205 |
1 files changed, 205 insertions, 0 deletions
diff --git a/app/src/lib/llm-service.ts b/app/src/lib/llm-service.ts new file mode 100644 index 0000000..af82fe2 --- /dev/null +++ b/app/src/lib/llm-service.ts @@ -0,0 +1,205 @@ +// import OpenAI from "openai"; +import { createUserContent, GoogleGenAI } from "@google/genai"; + +import type { + CategorizationRequest, + CategorizationResponse, +} from "./categorization"; +import { CATEGORIZATION_PROMPT } from "./llm-prompts"; + +export interface LLMRequestOptions { + model?: string; + temperature?: number; + maxTokens?: number; +} + +export class LLMService { + private defaultOptions: LLMRequestOptions = {}; + api; + model; + constructor( + // baseURL: string, + apiKey: string, + options: LLMRequestOptions = {}, + ) { + // const client = new OpenAI({ baseURL, apiKey }); + const client = new GoogleGenAI({ apiKey }); + this.api = client; + this.model = "gemini-2.5-flash"; + // this.defaultOptions = { + // model: "gemini-2.5-flash", + // temperature: 0.3, + // maxTokens: 1000, + // ...options, + // ...options + // }; + } + + async categorizeBookmark( + request: CategorizationRequest, + options?: LLMRequestOptions, + ): Promise<CategorizationResponse> { + const mergedOptions = { ...this.defaultOptions, ...options }; + + const prompt = this.buildCategorizationPrompt(request); + const media = []; + const allPics = request.bookmark.media.pics; + if (request.bookmark.media.video.thumb) + allPics.push(request.bookmark.media.video.thumb); + for (const pic of allPics) { + const imgdata = await fetch(pic); + const imageArrayBuffer = await imgdata.arrayBuffer(); + const base64ImageData = Buffer.from(imageArrayBuffer).toString("base64"); + const mimeType = imgdata.headers.get("content-type") || "image/jpeg"; + const con = { inlineData: { mimeType, data: base64ImageData } }; + media.push(con); + } + + const contents = media + ? createUserContent([ + prompt[0], + `**Media**: The bookmark included the following images:`, + ...media, + prompt[1], + ]) + : [prompt[0] + "\n\n" + prompt[1]]; + + try { + const response = await this.api.models.generateContent({ + model: this.model, + contents, + config: { systemInstruction: CATEGORIZATION_PROMPT }, + }); + console.log("llm res", response); + + if (!response.text) { + throw new Error("No response content from LLM"); + } + + return this.parseCategorizationResponse(response.text); + } catch (error) { + console.error("Error in LLM categorization:", error); + throw error; + } + } + + private buildCategorizationPrompt( + request: CategorizationRequest, + ): [string, string] { + const { bookmark, userCategories } = request; + + const prompt: [string, string] = [ + ` +Analyze the following bookmark and provide your categorization suggestions. + +## Bookmark Details +**Text**: "${bookmark.text}" +**Language**: ${bookmark.language} +**Author**: ${bookmark.author.name} (@${bookmark.author.username}) +**Hashtags**: ${bookmark.hashtags.join(", ") || "None"} +**URLs**: ${bookmark.urls.map((u) => u.expandedUrl).join(", ") || "None"}`, + + `## User Categories +${userCategories.map((cat) => `- ${cat.name}: ${cat.criteria}`).join("\n")} +Please provide your categorization analysis in the requested JSON format. +`, + ]; + + return prompt; + } + + private parseCategorizationResponse(content: string): CategorizationResponse { + try { + // Try to extract JSON from the response + const jsonMatch = content.match(/\{[\s\S]*\}/); + if (!jsonMatch) { + throw new Error("No JSON found in response"); + } + + const parsed = JSON.parse(jsonMatch[0]); + + // Validate response structure + if ( + !parsed.suggestedCategories || + !Array.isArray(parsed.suggestedCategories) + ) { + throw new Error( + "Invalid response structure: missing suggestedCategories", + ); + } + + return { + suggestedCategories: parsed.suggestedCategories.map((s: any) => ({ + categories: Array.isArray(s.categories) + ? s.categories + : [s.categories], + confidence: typeof s.confidence === "number" ? s.confidence : 0.5, + reasoning: s.reasoning || "No reasoning provided", + })), + newCategories: Array.isArray(parsed.newCategories) + ? parsed.newCategories + : [], + summary: parsed.summary || "No summary provided", + keyTopics: Array.isArray(parsed.keyTopics) ? parsed.keyTopics : [], + }; + } catch (error) { + console.error("Error parsing LLM response:", error); + + // Return fallback response + return { + suggestedCategories: [ + { + categories: ["Uncategorized"], + confidence: 0.1, + reasoning: "Failed to parse LLM response", + }, + ], + newCategories: [], + summary: "Analysis failed", + keyTopics: [], + }; + } + } + + // Generic LLM request method for future use + // async sendPrompt( + // prompt: string, + // options?: LLMRequestOptions, + // ): Promise<string> { + // const mergedOptions = { ...this.defaultOptions, ...options }; + + // try { + // const response = await fetch(`${this.baseUrl}/v1/messages`, { + // method: "POST", + // headers: { + // "Content-Type": "application/json", + // Authorization: `Bearer ${this.apiKey}`, + // "anthropic-version": "2023-06-01", + // }, + // body: JSON.stringify({ + // model: mergedOptions.model, + // max_tokens: mergedOptions.maxTokens, + // temperature: mergedOptions.temperature, + // messages: [ + // { + // role: "user", + // content: prompt, + // }, + // ], + // }), + // }); + + // if (!response.ok) { + // throw new Error( + // `LLM API request failed: ${response.status} ${response.statusText}`, + // ); + // } + + // const data = await response.json(); + // return data.content[0]?.text || ""; + // } catch (error) { + // console.error("Error in generic LLM request:", error); + // throw error; + // } + // } +} |