ludmi/dist/lu/retriever.js

LU (Layer Understanding) is a lightweight framework for controlled chatbot interactions with LLMs, action orchestration, and retrieval-augmented generation (RAG).

"use strict";
Object.defineProperty(exports, "__esModule", { value: true });
exports.rag = exports.retriever = void 0;
const openai_1 = require("./llm/openai");
//get the cosine similarity between two vectors
const cosineSimilarity = (vecA, vecB) => {
    const dotProduct = vecA.reduce((acc, cur, idx) => acc + cur * vecB[idx], 0);
    const normA = Math.sqrt(vecA.reduce((acc, cur) => acc + cur * cur, 0));
    const normB = Math.sqrt(vecB.reduce((acc, cur) => acc + cur * cur, 0));
    return dotProduct / (normA * normB);
};
//get chunks with the highest cosine similarity to the revised prompt
const retriever = async ({ revised_prompt, data, size = 4 }) => {
    const embedding = await (0, openai_1.getEmbeddings)(revised_prompt);
    const chunks = data.map(item => ({
        ...item,
        score: cosineSimilarity(embedding, item.embedding)
    }));
    chunks.sort((a, b) => b.score - a.score);
    return chunks.slice(0, size);
};
exports.retriever = retriever;
const rag = async ({ chunks, conversation, revised_prompt, userData = {}, model = "gpt-4o-mini", prompt }) => {
    //conversation o revised_promp, uno debe venir
    if (!conversation?.length && !revised_prompt)
        throw new Error("Debe enviar revised_prompt o conversation");
    const developerInstruction = prompt || `Responder solo usando la información extraída de la BD:
\`${JSON.stringify(chunks)}\`

REGLAS DE RESPUESTA:
- Mismo idioma del usuario.
- Estilo whatsapp, ejemplo para destacar palabra clave: *palabra* (No usar doble asteriscos; **MAL**).
- Emojis con coherencia.

Datos que conocemos del usuario: \`${JSON.stringify(userData)}\`
`;
    const messages = [
        {
            role: "developer",
            content: developerInstruction
        }
    ];
    if (revised_prompt)
        messages.push({
            role: "user",
            content: revised_prompt
        });
    else if (conversation)
        messages.push(...conversation);
    const AIresponse = await (0, openai_1.getAIResponse)({ messages, model, temperature: 1 });
    return AIresponse;
};
exports.rag = rag;