Pinecone RAG export

Record shape

Dataset-item exports use the item id, values: [], and metadata. The text to embed is metadata.text, taken from embedding_text, then context, then the source chunk, then the answer. Metadata also includes question, answer, source document id, dataset id, page, tags, difficulty, confidence, question type, and created time.

RAG Chunk exports use the same empty values array. Metadata then includes text from embedding_text, plus content, section, topic, keywords, entities, page, and quality fields.

Illustrative record

Synthetic. The empty values array is what the exporter writes.

{
  "id": "item-id",
  "values": [],
  "metadata": {
    "text": "Items can be returned within 30 days of delivery.",
    "question": "What is the return window?",
    "answer": "30 days from delivery."
  }
}

What you do next

1. Export format pinecone on a paid plan. Free export is JSON only.
2. Embed metadata.text with your model. Do not upsert the empty values array.
3. Upsert into your Pinecone index. FAQai does not host that index.
4. Query the index with the same embedding model and compare retrieved text with the source chunk.

Other vector targets are Qdrant, pgvector, Chroma, Weaviate, Milvus, LanceDB, and Upstash. Upstash is the exception that puts the text in data so Upstash can embed on import. Details are in the usage guide.