What is context precision vs context recall in RAG?

Beginner · 2 min read · Evaluation & Testing

Open in study mode to save this question and track progress

Short answer: Context Precision measures the signal-to-noise ratio of retrieved chunks (how many of the retrieved passages were actually relevant); Context Recall measures whether all the ground-truth information necessary to answer the prompt was successfully retrieved.

Detail

Evaluating the retrieval stage of a RAG pipeline requires separating retrieval precision from generation quality (e.g. in the Ragas evaluation framework).

Retrieved Context Chunks: [Chunk 1 (Relevant), Chunk 2 (Irrelevant), Chunk 3 (Irrelevant)]
  ├── Context Precision = Low (1/3 relevant chunks = high noise)
  └── Context Recall    = High (if Chunk 1 contains ALL necessary facts)

Metric	Focus Question	Target Metric Optimization
Context Precision	"Are top-ranked chunks actually relevant, or did we retrieve clutter?"	Optimizes rank position (Rerankers, higher similarity thresholds)
Context Recall	"Did the retriever find all key facts required to answer the question?"	Optimizes recall search (Increasing $K$ , hybrid search BM25 + Dense)

Context Precision Formula (Ragas)

$\text{Context Precision@K} = \frac{\sum_{k=1}^K \text{Precision@k} \times v_k}{\text{Total Relevant Chunks in Top K}}$

Where $v_k \in \{0, 1\}$ indicates if chunk $k$ is relevant.

Example

Python concept evaluating context recall against ground truth statements:

python

def calculate_context_recall(ground_truth_statements: list[str], retrieved_context: str) -> float:
    # Check what fraction of required ground truth facts exist in retrieved context
    found = 0
    for stmt in ground_truth_statements:
        if stmt.lower() in retrieved_context.lower():
            found += 1
    return found / len(ground_truth_statements) if ground_truth_statements else 0.0

gt = ["Server restarted at 02:00 UTC", "Error code was 503"]
context = "System logs show server restarted at 02:00 UTC due to maintenance."
print("Context Recall:", calculate_context_recall(gt, context)) # 0.5 (1 of 2 facts retrieved)

Interview tips

Emphasize that high context recall prevents hallucinations, while high context precision reduces prompt token costs and latency.
Connect this to the Ragas evaluation trilemma (Faithfulness, Context Recall, Answer Relevance).

⬅ Back to Evaluation and Testing · All topics