Ask the corpus three ways
The same question runs through RAG, GraphRAG and Agentic GraphRAG on the same TigerGraph graph. Every answer carries its evidence, its cost, and the reason the system stopped.
Inspect the investigation
Follow an answer from its wording to the graph operation and the original source span.
Pick a recorded prediction, or click “Inspect” on a live result.
Benchmark dashboard
Accuracy, completeness and token cost for every pipeline on the 100 public questions, plus the unscored hidden set and the labelled robustness set. Public results are scored locally (exact/numeric match and a blind LLM judge from a different model family); hidden accuracy is unknown to us.
No run selected.
Architecture
Corpus → TigerGraph (graph + vectors) → three pipelines → evidence report → offline evaluation.
Pipelines
RAG — one fixed similarity search over chunk vectors stored in TigerGraph (vectorSearch), one grounded generation. No graph traversal, no complete-set counting.
GraphRAG — the question is interpreted once, then one fixed retrieval template runs installed GSQL queries (cohort aggregate, maximum, venue→event, PREVIOUS_EVENT edge, facts with source spans). It cannot replan after seeing missing evidence.
Agentic GraphRAG — the same tools, but an orchestrator keeps an explicit ledger of evidence requirements. A deterministic coverage check reports the gaps; the orchestrator (LLM) picks the next specialist action (entity linking, graph traversal, similarity search, document retrieval, aggregation, prose repair, disambiguation) until the requirements are satisfied or a budget stops it. Every step is traced with its execution kind and tokens.
Fixed-plan control — the agent's planner and tools with replanning disabled: isolates the value of adaptation.
Graph model (TigerGraph 4.2, Savanna)
Document —HAS_CHUNK→ Chunk(embedding VECTOR 1536)
Document —DESCRIBES_EVENT→ Event —IN_EDITION→ Edition —PREVIOUS_EDITION→ Edition
Event —IN_SPORT→ Sport · Event —AT_VENUE→ Venue
Event —HAS_FACT→ Fact —SUPPORTED_BY→ Chunk (char offsets)
Entity —WON_MEDAL{medal,noc}→ Event · Entity —REPRESENTS→ Entity(NOC)
Event —PREVIOUS_EVENT{source}→ Event
Counts and superlatives are computed inside the graph over the complete corpus-defined cohort (sport × edition) and return the full candidate manifest with included / excluded / unknown status. A missing number is never treated as zero.
Original contribution
Coverage-directed investigation: the answer's evidence requirements are explicit objects; unresolved requirements drive the next action; evidence repair recovers missing infobox values from article prose with a verbatim, character-checked quote; ambiguity is reported instead of guessed.