landing_page/qdrant-landing/content/documentation/tutorials-search-engineering/retrieval-quality-golden-set.md

title, weight, aliases

title	weight	aliases
Building a Golden Query Set	6	/documentation/tutorials/retrieval-quality-golden-set/

Building a Golden Query Set

Time: 40 min	Level: Intermediate

This tutorial focuses on retrieval relevance: how well retrieved results match real user intent. To measure retrieval relevance, you need a labeled dataset of queries paired with their expected relevant documents (commonly called a golden query set or ground truth). This tutorial covers both building that dataset and running it through Qdrant to compute relevance metrics.

For orientation on the four layers of retrieval evaluation and where this tutorial fits, see Measuring ANN Precision.

Prerequisites. A Qdrant collection with your corpus indexed, an embedding model available to encode queries at evaluation time, and Python with ranx installed.

Generating Queries

There are three practical approaches to building a golden set. Each one trades quality against cost and scale.

1. Human Annotation

Domain experts assign relevance scores on a binary (relevant / not relevant) or graded (0/1/2 or 1–5) scale. Human-labeled data produces the highest-fidelity signal and is the primary source for graded labels. Expert time is the bottleneck, which typically limits this approach to a small set of high-value queries.

2. Real User Queries from Logs

Sample query-document pairs from your production logs, using clicks or explicit feedback (thumbs up/down, ratings) as the relevance signal. Real user queries capture intent and vocabulary that synthetic generation can't match, but you need enough traffic and a signal that maps to relevance.

Balance the sample so frequent queries don't crowd out rare ones: group by query type, topic, or intent class. Start with a few hundred labeled pairs to detect large metric differences; per-slice analysis or small ranking deltas need substantially more.

3. LLM-Based Synthetic Generation

An LLM can generate plausible queries for documents sampled from your corpus. This scales cheaply to thousands of pairs, but synthetic queries are typically easier to retrieve than real user queries, which inflates offline scores. For very large corpora, log-based sampling is often more practical.

The document you feed the LLM (the source document) becomes the relevance label for every query it generates:

You are helping build an evaluation dataset for a search system.

Generate 3 realistic search queries for the document below.
Each query should be what a real user would type to find it.
Phrase queries naturally, not as paraphrases of the document.
Return only the queries, one per line. No numbering or explanation.

Document:
{document_text}

Tune the prompt to your corpus:

• Query style. Questions for FAQ/RAG, keyword phrases for e-commerce, intent phrases for code search, or technical terms for specialist domains.
• Count per document. 3 is a default; tune to document length and golden-set size.
• Persona. A generic "user" works broadly; specialist corpora (medical, legal, technical) benefit from targeted personas.
• Language. Default English; state multilingual explicitly.

Using the Golden Set

ranx is a Python library for ranking-metric evaluation. It covers the standard IR metrics (recall@k, MRR, NDCG@k, Precision@k, MAP, and others) through one consistent interface, so you don't hand-roll each metric or juggle different libraries as needs grow.

The evaluation runs in three steps: load the labeled queries into the shape ranx expects, run each through Qdrant, then compute metrics.

1. Load and assemble. For each labeled query, build an entry with query_id, query_text, query_vector (embedded with the same model your Qdrant collection uses), and labels:

{
    "query_id": "q1",
    "query_text": "how does X work",
    "query_vector": [0.12, -0.48, 0.33, ...],  # embedding of query_text
    "labels": {"doc_42": 1},  # source doc for synthetic queries, relevant docs otherwise
}

Build the full golden_set by normalizing whatever your generation pipeline produced, then looping through it:

from your_embedding_model import embed

# Normalize whatever your generation pipeline produced into this shape:
#   - Synthetic: one item per generated query, labels = {source_doc_id: 1}
#   - Logs: one item per query-click pair, labels = {clicked_doc_id: 1}
#   - Human: one item per annotated query, labels = {doc_id: score, ...}
labeled_data = [
    {"query_text": "how does X work", "labels": {"doc_42": 1}},
    {"query_text": "what is Y used for", "labels": {"doc_55": 1, "doc_88": 1}},
    # ...one entry per labeled query
]

golden_set = []
for i, item in enumerate(labeled_data):
    golden_set.append({
        "query_id": f"q{i}",
        "query_text": item["query_text"],
        "query_vector": embed(item["query_text"]),
        "labels": item["labels"],
    })

2. Build Qrels and Run. ranx compares two inputs, both shaped as {query_id: {doc_id: score}}:

• Qrels (query relevance judgments). The labeled ground truth. Use 1 for binary labels or the raw 0/1/2 for graded labels.
• Run (retrieval output). What Qdrant returned for each query, with similarity scores.

from qdrant_client import QdrantClient
from ranx import Qrels, Run, evaluate

client = QdrantClient("http://localhost:6333")  # or QdrantClient(url="https://<id>.cloud.qdrant.io", api_key="...") for Qdrant Cloud

def retrieval_run(golden_set: list, collection: str, k: int = 10) -> Run:
    run = {}
    for entry in golden_set:
        results = client.query_points(
            collection_name=collection,
            query=entry["query_vector"],
            limit=k,
        ).points
        run[entry["query_id"]] = {p.id: p.score for p in results}
    return Run(run)

qrels = Qrels({entry["query_id"]: entry["labels"] for entry in golden_set})
run = retrieval_run(golden_set, collection="my_collection", k=10)

3. Compute metrics. evaluate(qrels, run, [...]) compares the two and returns a dict of metric names to floats.

metrics = evaluate(qrels, run, ["recall@10", "mrr", "ndcg@10"])

evaluate() returns:

{"recall@10": 0.82, "mrr": 0.71, "ndcg@10": 0.76}

Higher is better on all three. Which metric matters most depends on what your pipeline does with results:

Scenario	Recommended Metric	Why
RAG pipeline (LLM reads top-k chunks)	Recall@k	The LLM can recover if a relevant doc is at position 3 vs 1; missing it entirely hurts more
Single-answer retrieval (FAQ or Q&A)	MRR or Hits@1	The first result is what the user acts on; lower ranks matter little
Re-ranking or recommendation feeds	NDCG@k	Order within the result list matters; a highly relevant doc at rank 5 is worse than at rank 1

NDCG (Normalized Discounted Cumulative Gain) needs graded labels (for example, 0/1/2 scores per query-document pair). For binary labels, stick with recall@k and MRR. For the full metric list (Precision@k, MAP, ERR, and others), see the ranx docs.

On choosing k: set it to match actual usage. If the application shows 5 results to the user, measure @5. If a RAG pipeline passes 10 chunks to the LLM, measure @10. Reporting @100 for a UI that surfaces 5 results makes the metric look artificially good.

Re-run whenever the retrieval stack changes: new embedding model (which also requires re-embedding queries and re-indexing), new index config, or new reranker. In CI, compute recall@10 against a fixed golden set and fail the job when the score drops below your target threshold.

Pitfalls to Watch For (Data Leakage and Friends)

In golden query sets, data leakage means any setup that makes offline metrics look better than production reality. Unlike classic train/test leakage, the issue is often evaluation design. Keep source documents in the index (they are the expected relevant answers). Focus on these risks:

Synthetic-query unrealism. LLMs often mirror source wording, creating easier queries than real user input. This inflates offline scores. Mitigate it by instructing the LLM to generate queries as a user who hasn't seen the source document, then compare synthetic and real-query distributions (length and specificity).

Embedding-model contamination. If your embedding model was trained on pairs overlapping with the golden set, results will look better than true generalization. For hosted models, review published training data when possible. For in-house fine-tuning, keep strict train/eval separation.

Near-duplicate documents. Your retrieval may return a near-duplicate of a labeled document that isn't in the label set. That makes metrics look worse because labels are incomplete, not because retrieval is failing. A score dip here is a signal to audit your labels before tuning retrieval. Deduplicate before labeling (for example, cosine similarity > 0.95), or label duplicate clusters together.

Temporal drift. If the corpus changes after labeling, labels go stale: referenced docs may be removed or superseded by newer versions. Pin a corpus snapshot for each run and regenerate the golden set after material corpus changes.

Setup reproducibility. Version the full evaluation setup: corpus snapshot, how labels were produced, and any preprocessing thresholds. Otherwise you can't tell whether a later score drop is model/index regression or dataset drift.

Next Steps

Once retrieval relevance is on target, the next layer is pipeline output quality: whether the full pipeline produces the right output when retrieval feeds into a consumer (LLM generator, ranker, or UI). See Evaluating Pipeline Output Quality.