Audit GraphRAG relationships back to source passages
Trace graph edges to exact source passages, checking direction, entity identity, time, negation, and generated relationship descriptions.
Research briefing · Sources includedResearch & field notes
Research library — page 6
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.
Test GraphRAG community summaries for support and coverage
Evaluate generated community reports against graph facts and original passages, preserving minority exceptions, entity identity, and source support.
Research briefing · Sources includedTest GraphRAG index updates after a relationship changes
Test relationship edits, removals, entity changes, and affected summaries across a GraphRAG index without assuming a source update refreshes every artifact.
Research briefing · Sources includedTest RAG fallback search when the first retrieval fails
Evaluate fallback retrieval with explicit triggers, eligible sources, stopping rules, preserved user intent, and clear handling when recovery fails.
Research briefing · Sources includedDesign and test a useful zero-result RAG response
Distinguish genuine corpus absence, restrictive filters, ambiguous queries, and retrieval outages before designing a zero-result response.
Research briefing · Sources includedWrite RAG reference answers that preserve acceptable alternatives
Write evidence-backed reference answers with required facts, acceptable variations, scope, and explicit treatment of partially answerable questions.
Research briefing · Sources includedGive a RAG evaluator the right evidence without leaking labels
Give evaluators the actual retrieved evidence and task rules while keeping gold labels, unavailable documents, and hidden answer cues out of inappropriate inputs.
Research briefing · Sources includedClassify RAG questions before choosing retrieval metrics
Create a task-oriented query taxonomy for lookups, comparisons, multi-source reasoning, temporal questions, and unsupported requests.
Research briefing · Sources includedTurn retrieval errors into specific indexing and ranking fixes
Trace a failed answer through corpus eligibility, parsing, retrieval, reranking, and context assembly before choosing the next intervention.
Research briefing · Sources includedTest agent checkpoints before relying on recovery
Test agent checkpoint recovery around tool commits, restarts, schema changes, and approvals using a controlled ledger and explicit recovery criteria.
Research briefing · Sources includedTest agent idempotency when a tool call is repeated
Evaluate repeated agent tool calls with stable operation keys, conflicting arguments, concurrent retries, and an authoritative side-effect ledger.
Research briefing · Sources includedTest cancellation through an AI agent workflow
Test cancellation across queued tasks, model calls, tools, and late results so a stopped agent cannot silently continue consequential work.
Research briefing · Sources includedEvaluate agent recovery after a tool timeout
Evaluate agent timeout recovery by separating confirmed failures from unknown outcomes and testing reconciliation, fallback, and user-facing status.
Research briefing · Sources includedTest concurrent agent runs that touch shared state
Test concurrent agents against shared inventory, stale versions, duplicate tasks, and out-of-order results using a controlled operation ledger.
Research briefing · Sources includedTest agent approval boundaries before consequential actions
Test agent approvals against changed arguments, stale records, revoked authority, repeated resume requests, and downstream execution receipts.
Research briefing · Sources includedScope agent tool permissions with task-specific access
Scope agent tools to concrete task needs and test forbidden records, methods, destinations, and expired authority with permitted controls.
Research briefing · Sources includedClassify tool errors so an agent can recover appropriately
Classify agent tool failures by input, authorization, transient capacity, missing state, and uncertain writes before choosing a recovery action.
Research briefing · Sources includedDetect agent loops with observable stopping rules
Detect agent loops using repeated evidence, unchanged state, attempt budgets, and useful clarification rather than identical text alone.
Research briefing · Sources includedTest state consistency across agent steps
Test agent state consistency with versioned reads, delayed tool results, corrections, and write preconditions using a small controlled record.
Research briefing · Sources includedTest handoffs between agents with an explicit task contract
Evaluate agent-to-agent handoffs with explicit ownership, evidence references, permitted actions, deadlines, and acknowledgment of unresolved work.
Research briefing · Sources included