Reranking + Evaluation

What it is

Cohere Rerank for two-stage retrieval (broad ANN → precise rerank). Ragas + TruLens for measuring RAG quality (faithfulness, answer relevance, context precision) — so you actually know your system works.

How Vaaani uses it

• Reranking top-100 ANN hits down to top-5 with much higher precision
• Continuous eval pipelines that flag regressions before deploy
• A/B testing different retrieval strategies with objective scores
• Reporting accuracy metrics back to the business stakeholder

Why it makes the cut

Without eval, every prompt change is a gut feeling. With Ragas, every change has a number. Vaaani never ships RAG without it.

Sample code

from ragas import evaluate
from ragas.metrics import faithfulness, answer_relevancy

result = evaluate(
    dataset=eval_set,
    metrics=[faithfulness, answer_relevancy])
print(result)