Training Data API

The query log captures every request and response processed by Xilos, along with quality scores and routing metadata. Use these endpoints to export that data as a filtered training dataset for fine-tuning.

Endpoints

Method	Path	Description
`POST`	`/training-data/export`	Export filtered training data as a downloadable file
`GET`	`/training-data/preview`	Preview filtered training data before exporting

Export Training Data

POST /api/v1/training-data/export

Exports query-response pairs from the query log as a training dataset, filtered by quality scores, routing rule, and format.

Request Body

Field	Type	Required	Description
`format`	string	No	Output format: `jsonl`, `csv`, or `json`. Default: `jsonl`.
`min_faithfulness`	number	No	Minimum faithfulness score (0.0–1.0). Only includes responses with scores at or above this threshold. Default: `0.0`.
`min_relevance`	number	No	Minimum relevance score (0.0–1.0). Only includes responses with scores at or above this threshold. Default: `0.0`.
`routing_rule_id`	string	No	Only include queries that matched this routing rule.
`limit`	integer	No	Maximum number of records to export. Default: `1000`.

Info: Setting min_faithfulness and min_relevance to 0.7 or higher is recommended for fine-tuning. This ensures your training data only includes high-quality responses.

cURL

curl -X POST https://api.xilos.ai/api/v1/training-data/export \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_X..._KEY" \
  -d '{
    "format": "jsonl",
    "min_faithfulness": 0.8,
    "min_relevance": 0.8,
    "routing_rule_id": "rule_01HXYZ",
    "limit": 500
  }' \
  -o training_data.jsonl

Response

Returns a binary file download with the requested format. The response includes the following headers:

Header	Description
`Content-Type`	`application/jsonl`, `text/csv`, or `application/json`
`Content-Disposition`	`attachment; filename="training_data.jsonl"`
`X-Record-Count`	Number of records in the export

Preview Training Data

GET /api/v1/training-data/preview

Returns a preview of the filtered training data. Use this to inspect records before exporting the full dataset.

Query Parameters

Parameter	Type	Required	Description
`format`	string	No	Preview format: `jsonl`, `csv`, or `json`. Default: `jsonl`.
`min_faithfulness`	number	No	Minimum faithfulness score (0.0–1.0).
`min_relevance`	number	No	Minimum relevance score (0.0–1.0).
`routing_rule_id`	string	No	Filter by routing rule.
`limit`	integer	No	Max records to preview. Default: `10`. Max: `50`.

cURL

curl "https://api.xilos.ai/api/v1/training-data/preview?min_faithfulness=0.8&min_relevance=0.8&limit=5" \
  -H "Authorization: Bearer YOUR_X..._KEY"

Response

{
  "total_matching": 1247,
  "preview_count": 5,
  "records": [
    {
      "query": "Write a Python function to reverse a string",
      "response": "def reverse_string(s): return s[::-1]",
      "faithfulness": 0.92,
      "relevance": 0.95,
      "routing_rule_id": "rule_01HXYZ",
      "llm_id": "gpt-4o",
      "timestamp": "2025-01-15T10:30:00Z"
    }
  ]
}

Building a Training Dataset

Preview the Data

Use the preview endpoint with your desired filters to inspect sample records and verify quality.

Set Quality Thresholds

Apply min_faithfulness and min_relevance filters to exclude low-quality responses from your training data.

Export

Export the filtered dataset in your preferred format. JSONL is recommended for fine-tuning.

Create a Fine-Tune Job

Use the exported data to create a fine-tune job and produce a custom model.

Eval Datasets API Fine-Tune API