116 lines
3.5 KiB
JavaScript
116 lines
3.5 KiB
JavaScript
import Anthropic from "@anthropic-ai/sdk";
|
|
import { pipeline } from "node:stream/promises";
|
|
import { Transform, Writable } from "node:stream";
|
|
import { embedNotes } from "./embedding.service.js";
|
|
import { validateStructure, computeCohesionScore } from "./validation.service.js";
|
|
|
|
const client = new Anthropic({
|
|
apiKey: process.env.ANTHROPIC_API_KEY,
|
|
});
|
|
|
|
const buildPrompt = (notes) => {
|
|
const notesJson = JSON.stringify(notes, null, 2);
|
|
|
|
return `You are an expert at analyzing text for semantic similarity and thematic patterns.
|
|
|
|
Below is a JSON array of sticky notes. Each note has an "id" and a "text" field. Analyze the "text" field of every note and group them into meaningful thematic clusters.
|
|
|
|
For each cluster, return ONLY a valid JSON array with the below exact structure — no markdown, no explanation, no extra text - where the value for the "label" key is a name you create to describe the cluster's theme and the value for the "noteIds" key is an array containing the Ids of the notes that fit into that cluster theme.
|
|
|
|
[
|
|
{
|
|
"label": "Short descriptive theme name for cluster",
|
|
"noteIds": ["note_001", "note_002"]
|
|
}
|
|
]
|
|
|
|
Rules:
|
|
- Every note must appear in exactly one cluster
|
|
- Each cluster must have a concise, descriptive label
|
|
- Group by semantic meaning, not by keywords
|
|
- Aim for the most natural number of groups given the data
|
|
|
|
Here are the notes:
|
|
|
|
${notesJson}`;
|
|
};
|
|
|
|
const textDeltas = () => new Transform({
|
|
objectMode: true,
|
|
transform(event, _encoding, callback) {
|
|
if (event?.type === 'content_block_delta' && event.delta?.type === 'text_delta') {
|
|
callback(null, event.delta.text);
|
|
return;
|
|
}
|
|
callback();
|
|
},
|
|
});
|
|
|
|
// Rejects as soon as the first non-whitespace character proves the response
|
|
// is not the JSON array we asked for, rather than after the full generation.
|
|
const collectClusterJson = (sink) => new Writable({
|
|
objectMode: true,
|
|
write(text, _encoding, callback) {
|
|
if (!sink.sawOpeningBracket) {
|
|
const leading = (sink.parts.join('') + text).trimStart();
|
|
if (leading.length > 0) {
|
|
if (!leading.startsWith('[')) {
|
|
callback(new Error('LLM API returned non-JSON response'));
|
|
return;
|
|
}
|
|
sink.sawOpeningBracket = true;
|
|
}
|
|
}
|
|
sink.parts.push(text);
|
|
callback();
|
|
},
|
|
});
|
|
|
|
const requestClusters = async (notes, signal) => {
|
|
const sink = { parts: [], sawOpeningBracket: false };
|
|
|
|
const options = signal ? [{ signal }] : [];
|
|
|
|
const events = client.messages.stream(
|
|
{
|
|
model: "claude-sonnet-5",
|
|
max_tokens: 4096,
|
|
messages: [
|
|
{ role: "user", content: buildPrompt(notes) },
|
|
],
|
|
},
|
|
...options
|
|
);
|
|
|
|
await pipeline(events, textDeltas(), collectClusterJson(sink), ...options);
|
|
|
|
const text = sink.parts.join('');
|
|
|
|
if (text.trim().length === 0) {
|
|
throw new Error('Unexpected response from LLM API: no text content returned');
|
|
}
|
|
|
|
try {
|
|
return JSON.parse(text);
|
|
} catch {
|
|
throw new Error('LLM API returned non-JSON response');
|
|
}
|
|
};
|
|
|
|
export const clusterNotes = async (notes, { signal } = {}) => {
|
|
const [clusters, embeddingMap] = await Promise.all([
|
|
requestClusters(notes, signal),
|
|
embedNotes(notes),
|
|
]);
|
|
|
|
const noteIds = notes.map((n) => n.id);
|
|
const { valid, reasons } = validateStructure(clusters, noteIds);
|
|
if (!valid) {
|
|
throw new Error(`Cluster validation failed: ${reasons.join('; ')}`);
|
|
}
|
|
|
|
const score = computeCohesionScore(clusters, embeddingMap);
|
|
|
|
return { clusters, score: Math.round(score * 100) / 100 };
|
|
};
|