GraphRAG-Bench Novel Answer Accuracy

import { typegraphInit } from '@typegraph-ai/sdk'

const typegraph = await typegraphInit({
  apiKey: process.env.TYPEGRAPH_API_KEY!,
  tenantId: process.env.TYPEGRAPH_TENANT_ID!,
})

const bucket = await typegraph.bucket.create({
  name: 'graphrag-bench-novel',
  indexDefaults: {
    chunkSize: 512,
    chunkOverlap: 64,
    graphExtraction: true,
    deduplicateBy: ['content'],
  }
})

console.log(bucket.id)

import { readFile } from 'node:fs/promises'
import { typegraphInit } from '@typegraph-ai/sdk'

type NovelDocument = {
  id: string
  name: string
  corpus: string
  text: string
  url?: string
  metadata?: Record<string, unknown>
}

const typegraph = await typegraphInit({
  apiKey: process.env.TYPEGRAPH_API_KEY!,
  tenantId: process.env.TYPEGRAPH_TENANT_ID!,
})

const bucketId = process.env.TYPEGRAPH_BUCKET_ID!
const corpus = JSON.parse(await readFile('./novels.json', 'utf8')) as NovelDocument[]

const byCorpus = new Map<string, NovelDocument[]>()

for (const documentRow of corpus) {
  byCorpus.set(documentRow.corpus, [...(byCorpus.get(documentRow.corpus) ?? []), documentRow])
}

for (const [corpusId, documents] of byCorpus) {
  await typegraph.document.ingest(
    documents.map((documentRow) => ({
      id: documentRow.id,
      name: documentRow.name,
      content: documentRow.text,
      url: documentRow.url,
      metadata: {
        ...(documentRow.metadata ?? {}),
        corpus: corpusId,
        stableDocumentId: documentRow.id,
      },
    })),
    {
      bucketId,
      graphExtraction: true,
      chunkSize: 512,
      chunkOverlap: 64,
    },
  )
}

const response = await typegraph.search(question.text, {
  buckets: [process.env.TYPEGRAPH_BUCKET_ID!],
  graph: process.env.TYPEGRAPH_GRAPH_ID ?? 'public',
  resources: ['documents', 'facts', 'entities'],
  limit: 12,
  weights: {
    semantic: 1,
    bm25: 0.7,
    graph: 0.5,
    recency: 0.3,
  },
  promptBuilder: {
    format: 'markdown',
    sections: ['chunks', 'facts'],
    includeAttributes: false,
  },
})

// Use response.prompt as the full answer-generation context.
console.log(response.prompt)

import { appendFile, readFile } from 'node:fs/promises'
import { typegraphInit } from '@typegraph-ai/sdk'

type Question = {
  id: string
  corpus: string
  text: string
  questionType: string
}

const typegraph = await typegraphInit({
  apiKey: process.env.TYPEGRAPH_API_KEY!,
  tenantId: process.env.TYPEGRAPH_TENANT_ID!,
})

const questions = JSON.parse(await readFile('./queries.json', 'utf8')) as Question[]

for (const question of questions) {
  const response = await typegraph.search(question.text, {
    buckets: [process.env.TYPEGRAPH_BUCKET_ID!],
    graph: process.env.TYPEGRAPH_GRAPH_ID ?? 'public',
    resources: ['documents', 'facts', 'entities'],
    limit: 12,
    weights: { semantic: 1, bm25: 0.7, graph: 0.5 },
    promptBuilder: { format: 'markdown', sections: ['chunks', 'facts'] },
  })

  const answer = await generateAnswer({
    question: question.text,
    context: response.prompt,
  })

  await appendFile(
    './results.jsonl',
    JSON.stringify({
      id: question.id,
      corpus: question.corpus,
      questionType: question.questionType,
      answer,
      prompt: response.prompt,
      retrieval: response.results,
    }) + '\n',
  )
}

---Role---
You are a helpful assistant responding to user queries.

---Goal---
Generate a direct, concise answer based strictly on the provided Context.
Answer only what the Question asks. Do not restate the Question, explain your reasoning, or add background details.
Use one sentence when possible. For multi-part or creative requests, use the shortest complete answer that satisfies the Question.
If asked to summarize, summarize the relationships, effect, contrast, or implication asked about, not the whole passage.
Stay grounded in the Context and avoid unsupported specifics.
If the Context contains partial relevant evidence, synthesize the supported answer instead of refusing.
When the question asks for a chain, connection, relation, sequence, or how X links to Y, answer with only the explicit named relation chain from the context; do not explain themes, causes, motives, or background unless asked.
Respond in plain text without formatting.
Use the same language as the Question.
No markdown headings, no placeholder dates, do not refuse historical/fictional perspective tasks.
Default to 5-20 words; exceed 25 words only when the Question explicitly asks for a summary, comparison, explanation, or creative response.
If the answer can be expressed as a name, list, date, place, relation, or short clause, output only that.