Test RAG metadata filters with explicit eligible documents
Build positive and negative filter cases for tenant, date, language, and document status without confusing eligibility with relevance.
Research briefing · Sources includedResearch & field notes
Research library — page 4
Read a guide, follow its sources, and check the limits before applying a finding. Search the full library from the research desk, or browse by topic.
Compare hybrid search with lexical and vector baselines
Compare lexical, vector, and hybrid retrieval using the same eligible corpus, candidate budget, query slices, and answer checks.
Research briefing · Sources includedEvaluate a RAG reranker without hiding candidate-retrieval failures
Evaluate reranking on a frozen candidate pool and end to end, preserving failures where the first retriever never found the evidence.
Research briefing · Sources includedEvaluate an embedding-model migration before reindexing
Compare embedding versions with paired queries, index lineage, language slices, and a reversible cutover to a separately built index.
Research briefing · Sources includedTest query rewriting for preserved intent and useful recall
Test rewritten questions for retained entities, exclusions, dates, and permissions before rewarding changes in retrieval recall.
Research briefing · Sources includedEvaluate multi-query retrieval against its extra cost
Compare multiple query variants with a single-query baseline using unique evidence, merged rankings, latency, and cost per accepted answer.
Research briefing · Sources includedTest parent-child retrieval for context and precise evidence
Test child matches and parent expansion separately, checking eligibility, evidence preservation, and the cost of larger returned context.
Research briefing · Sources includedEvaluate context compression for lost evidence
Evaluate compressed context against original passages, retaining qualifiers, contradictions, citations, and evidence needed for abstention.
Research briefing · Sources includedTest whether context order changes a RAG answer
Reorder identical retrieved evidence to test position sensitivity without changing source content, document eligibility, or question wording.
Research briefing · Sources includedUse mean reciprocal rank for RAG with explicit relevance labels
Calculate mean reciprocal rank with explicit relevance judgments, cutoff rules, missing-result handling, and limits for multi-source questions.
Research briefing · Sources includedUse nDCG to evaluate graded retrieval relevance
Use graded relevance labels, a declared gain formula, and a fixed cutoff to interpret nDCG without confusing ranking with answer quality.
Research briefing · Sources includedChoose and interpret precision at k for retrieval
Define the retrieval unit, cutoff, eligibility rules, and duplicate handling before using precision at k to compare RAG contexts.
Research briefing · Sources includedMeasure duplicate passages in a RAG context window
Detect exact and near-duplicate retrieved passages, preserving meaningful version differences and measuring evidence displaced from the context.
Research briefing · Sources includedMeasure citation precision and citation recall separately
Score citation support and claim coverage separately with an explicit claim inventory, source passages, and handling for compound statements.
Research briefing · Sources includedEvaluate RAG groundedness at the claim level
Check whether answer claims follow from supplied evidence, distinguishing groundedness from relevance, correctness, completeness, and source authority.
Research briefing · Sources includedCheck whether a RAG answer covers the required sources
Check whether answers cover each required evidence group rather than rewarding many citations to one source or one side of a question.
Research briefing · Sources includedBuild unanswerable questions that meaningfully test RAG
Create corpus-bounded unanswerable questions with near matches, missing constraints, and evidence audits that distinguish absence from retrieval failure.
Research briefing · Sources includedTest RAG answers to questions whose facts change over time
Evaluate questions with explicit as-of dates, effective periods, timezone boundaries, and stale sources before treating a recently indexed fact as current.
Research briefing · Sources includedVerify deleted documents stop influencing RAG answers
Test deleted-source behavior across search indexes, child chunks, derived summaries, retrieved-context caches, and answer citations.
Research briefing · Sources includedTest incremental indexing for missing and duplicate records
Use an expected-record manifest to test incremental adds, edits, deletes, retries, and interrupted runs without relying on job success alone.
Research briefing · Sources included