Spaces:
Sleeping
Sleeping
File size: 9,191 Bytes
097eaf0 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 | import { MapDocument, URLTrace } from "../../controllers/v1/types";
import { logger } from "../logger";
import { generateCompletions } from "../../scraper/scrapeURL/transformers/llmExtract";
import { buildRerankerUserPrompt } from "./build-prompts";
import { getModel } from "../generic-ai";
import { CostTracking } from "../cost-tracking";
const THRESHOLD_FOR_SINGLEPAGE = 0.6;
const THRESHOLD_FOR_MULTIENTITY = 0.45;
type RerankerResult = {
mapDocument: (MapDocument & { relevanceScore?: number; reason?: string })[];
tokensUsed: number;
cost: number;
};
type RerankerOptions = {
links: MapDocument[];
searchQuery: string;
urlTraces: URLTrace[];
isMultiEntity: boolean;
reasoning: string;
multiEntityKeys: string[];
keyIndicators: string[];
costTracking: CostTracking;
metadata: { teamId: string; functionId?: string; extractId?: string };
};
export async function rerankLinksWithLLM(
options: RerankerOptions,
): Promise<RerankerResult> {
const {
links,
searchQuery,
urlTraces,
isMultiEntity,
reasoning,
multiEntityKeys,
keyIndicators,
metadata,
} = options;
const chunkSize = 5000;
const chunks: MapDocument[][] = [];
const TIMEOUT_MS = 60000;
const MAX_RETRIES = 2;
let totalTokensUsed = 0;
// await fs.writeFile(
// `logs/links-${crypto.randomUUID()}.txt`,
// JSON.stringify(links, null, 2),
// );
// Split links into chunks of 200
for (let i = 0; i < links.length; i += chunkSize) {
chunks.push(links.slice(i, i + chunkSize));
}
// console.log(`Total links: ${mappedLinks.length}, Number of chunks: ${chunks.length}`);
const schema = {
type: "object",
properties: {
relevantLinks: {
type: "array",
items: {
type: "object",
properties: {
url: { type: "string" },
relevanceScore: { type: "number" },
reason: {
type: "string",
description:
"The reason why you chose the score for this link given the intent.",
},
},
required: ["url", "relevanceScore", "reason"],
},
},
},
required: ["relevantLinks"],
};
let totalCost = 0;
const results = await Promise.all(
chunks.map(async (chunk, chunkIndex) => {
// console.log(`Processing chunk ${chunkIndex + 1}/${chunks.length} with ${chunk.length} links`);
const linksContent = chunk
.map(
link =>
`URL: ${link.url}${link.title ? `\nTitle: ${link.title}` : ""}${link.description ? `\nDescription: ${link.description}` : ""}`,
)
.join("\n\n");
// fs.writeFile(
// `logs/links-content-${crypto.randomUUID()}.txt`,
// linksContent,
// );
for (let retry = 0; retry <= MAX_RETRIES; retry++) {
try {
// const timeoutPromise = new Promise<null>(resolve => {
// setTimeout(() => resolve(null), TIMEOUT_MS);
// });
const systemPrompt = `You are analyzing URLs for ${isMultiEntity ? "collecting multiple items" : "specific information"}.
The user's query is: ${searchQuery}
${
isMultiEntity
? `IMPORTANT: This is a multi-entity extraction task looking for ${multiEntityKeys.join(", ")}.
Score URLs higher if they contain ANY instance of the target entities.
Key indicators to look for: ${keyIndicators.join(", ")}`
: `IMPORTANT: This is a specific information task.
Score URLs based on precision and relevance to answering the query.`
}
Scoring guidelines:
${
isMultiEntity
? `
- 1.0: Contains ANY instance of target entities, even just one. Give this score if page has any relevant entity. If you are not sure if this page is relevant or not, give it a score of 1.0
- 0.8: Contains entity but may be incomplete information
- 0.6: Mentions entity type but no clear instance
- 0.4: Only tangentially related to entity type
- Below 0.4: No mention of relevant entities, or duplicates
Reason: ${reasoning}
`
: `
- 1.0: Contains direct, authoritative answer to query. Give this score if unsure about relevance. If you are not sure if this page is relevant or not, give it a score of 1.0
- 0.8: Contains information that directly helps answer the query
- 0.6: Contains related information that partially answers query
- Below 0.6: Information too general or not focused on query
`
}`;
// dumpToFile(new Date().toISOString(),[buildRerankerSystemPrompt(), buildRerankerUserPrompt(searchQuery), schema, linksContent])
// const gemini = getGemini();
// const model = getGemini()
let completion: any;
try {
const completionPromise = generateCompletions({
model: getModel("gemini-2.5-pro", "vertex"),
retryModel: getModel("gemini-2.5-pro", "google"),
logger: logger.child({
method: "rerankLinksWithLLM",
chunk: chunkIndex + 1,
retry,
}),
options: {
systemPrompt: systemPrompt,
prompt: buildRerankerUserPrompt(searchQuery),
schema: schema,
// temperature: isMultiEntity ? 0.5 : 0.3,
},
// providerOptions: {
// anthropic: {
// thinking: { type: 'enabled', budgetTokens: 12000 },
// tool_choice: "auto",
// },
// },
markdown: linksContent,
isExtractEndpoint: true,
costTrackingOptions: {
costTracking: options.costTracking,
metadata: {
module: "extract",
method: "rerankLinksWithLLM",
},
},
metadata: {
...metadata,
functionId: metadata.functionId
? metadata.functionId + "/rerankLinksWithLLM"
: "rerankLinksWithLLM",
},
});
completion = await completionPromise;
totalCost += completion.cost;
} catch (error) {
console.warn(
`Error processing chunk ${chunkIndex + 1} attempt ${retry + 1}:`,
error,
);
}
// await fs.writeFile(
// `logs/reranker-${crypto.randomUUID()}.json`,
// JSON.stringify(completion, null, 2),
// );
if (!completion) {
// console.log(`Chunk ${chunkIndex + 1}: Timeout on attempt ${retry + 1}`);
continue;
}
if (!completion.extract?.relevantLinks) {
// console.warn(`Chunk ${chunkIndex + 1}: No relevant links found in completion response`);
return [];
}
totalTokensUsed += completion.numTokens || 0;
// console.log(`Chunk ${chunkIndex + 1}: Found ${completion.extract.relevantLinks.length} relevant links`);
return completion.extract.relevantLinks;
} catch (error) {
console.warn(
`Error processing chunk ${chunkIndex + 1} attempt ${retry + 1}:`,
error,
);
if (retry === MAX_RETRIES) {
// console.log(`Chunk ${chunkIndex + 1}: Max retries reached, returning empty array`);
return [];
}
}
}
return [];
}),
);
// console.log(`Processed ${results.length} chunks`);
// Flatten results and sort by relevance score
const flattenedResults = results
.flat()
.sort((a, b) => b.relevanceScore - a.relevanceScore);
// console.log(`Total relevant links found: ${flattenedResults.length}`);
// Map back to MapDocument format, keeping ALL links for testing
const relevantLinks = flattenedResults
.map(result => {
if (
result.relevanceScore >
(isMultiEntity ? THRESHOLD_FOR_MULTIENTITY : THRESHOLD_FOR_SINGLEPAGE)
) {
const link = links.find(link => link.url === result.url);
if (link) {
return {
...link,
relevanceScore: result.relevanceScore
? parseFloat(result.relevanceScore)
: 0,
reason: result.reason,
};
}
}
return undefined;
})
.filter((link): link is NonNullable<typeof link> => link !== undefined);
// Add debug logging for testing
// fs.writeFile(
// `logs/reranker-aaa-${crypto.randomUUID()}.json`,
// JSON.stringify(
// {
// totalResults: relevantLinks.length,
// scores: relevantLinks.map((l) => ({
// url: l.url,
// score: l.relevanceScore,
// reason: l.reason,
// })),
// },
// null,
// 2,
// ),
// );
return {
mapDocument: relevantLinks,
tokensUsed: totalTokensUsed,
cost: totalCost,
};
}
|