Back to Docs
Benchmarks
CSV Upload
Import existing Q&A datasets from CSV files
Last updated: August 20, 2025
Category: benchmarks
CSV Upload
Import an existing question-answer dataset by uploading a CSV file. This is useful when you already have manually curated test cases or want to bring in datasets from other tools.
CSV Format
Required Columns
| Column | Description |
|---|---|
question | The query text |
answer | The expected answer |
Optional Columns
| Column | Format | Description |
|---|---|---|
id | String | Unique entry ID (auto-generated if omitted) |
chunk_ids | Pipe-separated | Ground-truth chunk IDs (e.g., chunk_1|chunk_2) |
page_nums | Pipe-separated integers | Page numbers (e.g., 1|2|3) |
source_paths | Pipe-separated | Document names (e.g., report.pdf|manual.pdf) |
Example CSV
question,answer,chunk_ids,page_nums,source_paths
"What is the return policy?","Items can be returned within 30 days.","chunk_12|chunk_15","1|2","returns_policy.pdf"
"How do I reset my password?","Go to Settings > Security > Reset Password.","chunk_42","5","user_guide.pdf"
"What are the shipping options?","Standard (5-7 days) and Express (1-2 days).","chunk_8|chunk_9|chunk_10","3|4","shipping_faq.pdf|logistics.pdf"
Important: Multi-value fields use the pipe character (
|) as a separator, not commas, because the file itself is comma-separated.
Uploading via the Platform
- Navigate to Benchmarks
- Click Upload CSV
- Provide a benchmark name
- Select your CSV file
- Click Upload
The benchmark is created with status active and is immediately ready for evaluation.
Uploading via the API
The CSV upload endpoint is available through the backend API:
POST /benchmarks/upload-csv
Content-Type: multipart/form-data
Fields:
- name: "My Benchmark"
- description: "Optional description"
- csv_file: <your_file.csv>
When to Use CSV Upload
- Manual curation — You have hand-crafted test cases from domain experts
- Migration — Moving evaluation data from another tool
- Generation-only benchmarks — You only need
questionandanswercolumns (nochunk_idsneeded) - External pipelines — Your Q&A pairs are produced by a separate script or system
Tips
- Benchmarks without
chunk_idscan still be used for generation-only evaluations - Benchmarks without
page_numswill skip page-level metrics (chunk and document metrics still work) - Exported benchmarks use the same CSV format, so you can download, edit, and re-upload
Next Steps
- Hugging Face — Import standard research datasets
- Synthetic Generation — Auto-generate from your data
- Evaluations — Run evaluations against your benchmark