Configuration Reference

Complete reference for RAGAS adapter configuration options.

JobSpec Structure

The RAGAS adapter uses a standardised JobSpec structure:

{
  "id": "string",
  "provider_id": "ragas",
  "benchmark_id": "string",
  "model": {
    "name": "string",
    "url": "string"
  },
  "parameters": {
    // RAGAS-specific configuration
  },
  "num_examples": 0
}

Core Parameters

Required Parameters

Parameter	Type	Description	Example
id	string	Unique job identifier	"ragas-rag-eval-001"
provider_id	string	Must be "ragas"	"ragas"
benchmark_id	string	Benchmark identifier	"ragas_rag_default"
model.name	string	Model name for the LLM judge	"Qwen/Qwen2.5-1.5B-Instruct"
model.url	string	OpenAI-compatible API endpoint	"http://localhost:8000"

Optional Parameters

Parameter	Type	Description	Default
num_examples	integer	Limit the number of dataset samples to evaluate	All samples
callback_url	string	EvalHub service callback URL	null

Benchmark Configuration

Two pre-defined benchmark suites are available:

ragas_rag_default: Default Suite

Runs the four core RAG evaluation metrics. Suitable for most use cases.

Setting	Value
Metrics	answer_relevancy, context_precision, faithfulness, context_recall
Primary score	faithfulness
Pass threshold	0.5

ragas_rag_full: Full Suite

Runs all 11 available metrics for comprehensive RAG evaluation.

Setting	Value
Metrics	All 11 metrics (see Metrics reference)
Primary score	faithfulness
Pass threshold	0.5

Parameters Reference

All configuration is specified in the parameters object of the JobSpec.

Metric Selection

Parameter	Type	Description	Default
metrics	array	List of RAGAS metric names to evaluate	Benchmark default

Available metrics: answer_relevancy, answer_similarity, context_precision, faithfulness, context_recall, context_entity_recall, nv_accuracy, nv_context_relevance, factual_correctness, noise_sensitivity, nv_response_groundedness.

See the Metrics reference for details on each metric.

LLM Configuration

Parameter	Type	Description	Default
max_tokens	integer	Maximum tokens for LLM completions	null (server default)
temperature	number	Sampling temperature for LLM completions	null (adapter default)

Embedding Configuration

Parameter	Type	Description	Default
embedding_model	string	Model name for embeddings	Same as model.name
embedding_url	string	Base URL for embeddings endpoint	Same as model.url

Embedding Model Selection

Several RAGAS metrics (notably answer_relevancy and answer_similarity) require embedding support. If your LLM endpoint does not serve embeddings, you must configure a separate embedding_url and embedding_model. A lightweight embedding model such as all-MiniLM-L6-v2 or nomic-embed-text is recommended.

Data Configuration

Parameter	Type	Description	Default
data_path	string	Explicit path to dataset file	Auto-resolved
column_map	object	Map dataset column names to RAGAS expected names	null

Column mapping: RAGAS expects columns named user_input, response, retrieved_contexts, and reference. If your dataset uses different names, provide a mapping:

{
  "parameters": {
    "column_map": {
      "question": "user_input",
      "answer": "response",
      "contexts": "retrieved_contexts",
      "ground_truth": "reference"
    }
  }
}

Data resolution order (when data_path is not set):

1. /test_data/dataset.jsonl, populated by EvalHub’s S3 init container
2. First .jsonl or .json file in /test_data/
3. /data/dataset.jsonl
4. First .jsonl or .json file in /data/

Concurrency

Parameter	Type	Description	Default
max_workers	integer	Parallel workers for RAGAS evaluation (1–10)	1

Concurrency Trade-off

Increasing max_workers speeds up evaluation but sends more concurrent requests to the judge endpoint. Start with 1 and increase if the judge endpoint has sufficient capacity.

LLM Judge Requirements

RAGAS uses an LLM as a judge for several metrics. The judge model receives structured prompts and must return parseable JSON responses.

Which Metrics Require a Judge

Metric	LLM Judge	Embeddings
faithfulness	Yes	No
answer_relevancy	Yes	Yes
context_precision	Yes	No
context_recall	Yes	No
answer_similarity	No	Yes
context_entity_recall	No	No
factual_correctness	Yes	No
noise_sensitivity	Yes	No
nv_accuracy	Yes	No
nv_context_relevance	Yes	No
nv_response_groundedness	Yes	No

Cost and Latency Implications

• Each sample is evaluated independently per metric. For N samples and M judge-based metrics, expect roughly N * M LLM calls.
• Judge prompts are structured and can be lengthy, so set max_tokens appropriately (512 is usually sufficient).
• Use a lower temperature (e.g. 0.1) for more deterministic judge outputs.
• The adapter uses chat completions (not legacy completions) to avoid truncation issues.

Environment Variables

The adapter reads runtime settings from environment variables:

Variable	Description	Required	Default
EVALHUB_MODE	Execution mode (k8s or local)	No	k8s
EVALHUB_JOB_SPEC_PATH	Path to job spec JSON	Yes (local mode)	/meta/job.json
SERVICE_URL	Eval-hub service URL	No	null
REGISTRY_URL	OCI registry URL	No	null
REGISTRY_USERNAME	Registry username	No	null
REGISTRY_PASSWORD	Registry password	No	null
REGISTRY_INSECURE	Allow insecure registry	No	false
LOG_LEVEL	Logging level	No	INFO

Complete Example

{
  "id": "ragas-rag-eval-001",
  "provider_id": "ragas",
  "benchmark_id": "ragas_rag_default",
  "benchmark_index": 0,
  "model": {
    "url": "http://127.0.0.1:8000",
    "name": "Qwen/Qwen2.5-1.5B-Instruct"
  },
  "num_examples": 5,
  "parameters": {
    "metrics": [
      "answer_relevancy",
      "context_precision",
      "faithfulness",
      "context_recall"
    ],
    "embedding_model": "all-MiniLM-L6-v2",
    "embedding_url": "http://127.0.0.1:8001",
    "max_tokens": 512,
    "temperature": 0.1
  },
  "callback_url": "http://localhost:8080"
}