hydrag-benchmark 0.5.3

Local-only RAG benchmarking CLI — measures recall, MRR, chunk overlap, latency, and BEIR IR metrics

hydrag-benchmark

Local-only RAG benchmarking CLI for retrieval quality and latency analysis.

Installation

pip install hydrag-benchmark

Optional GPU path for multi-head dense embeddings:

pip install "hydrag-benchmark[gpu]"

Included Suites

• suites/synthetic-smoke.yaml
• suites/k8s-kep.yaml
• suites/cpython-stdlib.yaml

Quickstart

# List shipped suites
hydrag-bench list-suites --suite-dir ./suites

# Run classic strategy benchmark
hydrag-bench run suites/synthetic-smoke.yaml \
  --strategy hydrag \
  --corpus-dir ./my-codebase/src \
  --output-dir ./results

# Inspect output
python -m json.tool ./results/synthetic-smoke_hydrag.json

Commands

hydrag-bench --help
hydrag-bench --version

# 1) Classic single-strategy benchmark
hydrag-bench run <suite.yaml> --strategy <similarity|hybrid|crag|hydrag> --corpus-dir <path> [options]

# 2) List suites
hydrag-bench list-suites --suite-dir <path>

# 3) Prefill Doc2Query cache (Phase 1a)
hydrag-bench prefill --corpus-dir <path> [options]

# 4) Multi-head harness benchmark (Heads A/B/C)
hydrag-bench multihead <suite.yaml> --corpus-dir <path> [options]

run Arguments

Flag	Required	Default	Description
suite	yes	-	Path to benchmark suite YAML
--strategy	yes	-	One of similarity, hybrid, crag, hydrag
--corpus-dir	yes	-	Root directory of files to index
--output-dir	no	stdout	Directory to write <suite>_<strategy>.json
--suite-dir	no	-	Base dir for resolving relative suite path
--n-results	no	5	Top-k retrieval depth
--seed	no	42	Seed override
--embedding-model	no	Alibaba-NLP/gte-Qwen2-7B-instruct	Embedding model label passed to runner
--db-path	no	temp dir	ChromaDB persistence path

list-suites Arguments

Flag	Required	Default	Description
--suite-dir	yes	-	Directory containing .yaml / .yml suites

prefill Arguments

Flag	Required	Default	Description
--corpus-dir	yes	-	Root directory to chunk and process
--doc2query-model	no	qwen3:4b	Doc2Query model name
--doc2query-api-url	no	http://localhost:11434	Doc2Query API base URL
--doc2query-timeout-s	no	30.0	Request timeout seconds
--doc2query-max-retries	no	2	Retry attempts after first failure
--doc2query-n-questions	no	3	Synthetic questions per chunk
--cache-dir	no	in-memory only	Directory containing augmentation_cache.json

multihead Arguments

Flag	Required	Default	Description
suite	yes	-	Path to benchmark suite YAML
--corpus-dir	yes	-	Root directory of files to index
--output-dir	no	stdout	Directory to write <suite>_multihead.json and sidecar
--suite-dir	no	-	Base dir for resolving relative suite path
--n-results	no	5	Top-k retrieval depth
--seed	no	42	Seed override
--use-gpu	no	false	Use transformers embedder (requires [gpu])
--doc2query-model	no	qwen3:4b	Doc2Query model name
--doc2query-api-url	no	http://localhost:11434	Doc2Query API base URL
--doc2query-timeout-s	no	30.0	Request timeout seconds
--doc2query-max-retries	no	2	Retry attempts after first failure
--doc2query-n-questions	no	3	Synthetic questions per chunk
--embedding-model	no	Alibaba-NLP/gte-Qwen2-7B-instruct	Dense embedding model name
--alpha	no	0.5	Head C rerank interpolation weight
--cache-dir	no	none	Directory for augmentation_cache.json persistence

Config Variables and Runtime Inputs

• hydrag-benchmark does not read HYDRAG_BENCHMARK_* environment variables.
• Operator-facing runtime configuration is via CLI flags and suite YAML fields.
• Suite-level fields consumed by code:
  • top-level: name, version, seed, description, cases
  • environment: strategy, n_results

File Paths and Artifacts

Path / Pattern	Producer	Meaning
<output-dir>/<suite>_<strategy>.json	run	Single-strategy result JSON (schema_version: 0.1)
<output-dir>/<suite>_multihead.json	multihead	Multi-head comparison matrix (schema_version: 0.2)
<output-dir>/questions_sidecar.json	multihead	Head B generated questions sidecar
<cache-dir>/augmentation_cache.json	prefill / multihead	3-state Doc2Query cache shared across phases
<db-path>	run	ChromaDB persistent store location

Output Schemas

• run emits schema 0.1 with per-case and aggregate metrics.
• multihead emits schema 0.2 with 5 config groups:
  • A-only
  • B-only
  • C-only
  • A+B
  • A+B+C

Frozen 0.1 Metrics

Metric	Description
recall_at_1	1.0 when top result includes a relevant phrase
recall_at_k	Fraction of relevant phrases found in top-k
mrr	Mean Reciprocal Rank of first relevant result
chunk_overlap	Token overlap between retrieved chunks and relevant phrases
latency_ms.avg	Mean latency in milliseconds
latency_ms.p50	50th percentile latency
latency_ms.p95	95th percentile latency
latency_ms.p99	99th percentile latency

Suite YAML Format

name: my-benchmark
version: "1.0"
seed: 42
description: Description of the benchmark suite.

environment:
  strategy: hydrag
  n_results: 5

cases:
  - id: case-001
    query: "search query text"
    relevant_phrases:
      - "expected phrase in results"
      - "another expected phrase"
    tags: [optional, tags]

Development

cd packages/hydrag-benchmark
pip install -e ".[dev]"
python -m pytest tests/ -v

License

Apache-2.0