Waiting to extract the complete event
Keep roles, schedule changes and conditions together for retrieval judgmentsSAG
Find the evidence an agent needs across documents.
Find the facts an answer needs, across documents
An answer may depend on facts spread across several documents. SAG records each passage's events and the people, companies and projects involved. At query time, it follows those connections to retrieve the supporting text.
- Keep each event intact
- Record what a passage says happened, along with the people, companies and projects it mentions.
- Follow connections across documents
- Find a project's lead, then use that person's name to find team information missing from the first search.
- Return the original text
- Give agents the passages needed to answer, and let users check each source.
90.07%Evidence found in the top fiveThree-benchmark mean · Recall@5
72.96Answer match scoreThree-benchmark mean · F1 (0–100)
+11.52MuSiQue recall gainPoints vs strongest paper baseline
Results from the papers. Benchmarks, metrics and test conditions are explained below.
01 / HOW IT WORKS
Extract events and entities, then query their connections
SAG prepares separate event and entity records from each source passage, then stores their associations. A question activates shared-entity joins, and results map back to the original text.
EVENT–ENTITY ASSOCIATIONS1 / 6
Which team does Project Aurora's lead belong to?
PREPARATION
Source → parallel event and entity extraction → associations
Events and entities are two outputs extracted from the same sourceSOURCE A · project.md · passage 3
Friday's review began with development progress on Aurora Enterprise. Only internal trials are complete, and single sign-on integration is still unfinished, so the product cannot be released publicly yet. The original May launch must therefore change. The meeting agreed to move the release to June, conditional on completing single sign-on integration and passing the security review. Lin Ran owns this release and coordinates testing, acceptance and launch arrangements. Zhou Ning remains responsible for performance testing and does not make release decisions. External notices will be sent after acceptance. The April demo was for internal use only and did not constitute a public release.
LLM extracts both from the source
Waiting to extract source entities
Extract from the source to index events and connect passages mentioning the same entitiesEvent A ↔ Aurora Enterprise, Lin Ran, Zhou Ning, April, May, June, single sign-on, security acceptance
One source passage maps to one complete event, associated with the entities extracted from that sourceKeep the source passages
The source mixes progress, a delay, two people's roles and several dates. Retain the passage and its location so answers can be checked against it.
A fictional example illustrating the retrieval process. Select any step to inspect it; autoplay stops after one pass.
VIDEO / SAG
SAG explained
Follow SAG from event and entity extraction to retrieving connected evidence across documents.
Video and resources on GitHub02 / BENCHMARKS
Find more evidence and give better answers
Recall@5 measures how much required evidence appears in the first five results. F1 measures how closely the answer matches the reference. Switch benchmarks to compare retrieval and answers separately.
MuSiQueRecall@5 · Higher is better
+11.52 points vs NV-Embed-v2
SAG80.36%
NV-Embed-v268.84%
GritLM-7B66.22%
HippoRAG 265.13%
GTE-Qwen2-7B61.92%
BGE-Large-EN-v1.556.32%
Contriever45.32%
BM2542.11%
1,000 development questions per benchmark. All methods use Qwen3.6-Flash to answer; structure-augmented methods use BGE-Large-EN-v1.5 for retrieval. Methods with no reported Recall are omitted from that chart.
Full benchmark dataRecall@5
| Method | MuSiQue | 2WikiMultiHopQA | HotpotQA |
|---|---|---|---|
| BM25 | 42.11 | 63.78 | 73.05 |
| Contriever | 45.32 | 59.25 | 74.85 |
| BGE-Large-EN-v1.5 | 56.32 | 69.83 | 89.20 |
| GTE-Qwen2-7B | 61.92 | 74.18 | 88.45 |
| GritLM-7B | 66.22 | 75.62 | 93.40 |
| NV-Embed-v2 | 68.84 | 77.00 | 95.25 |
| GraphRAG | — | — | — |
| LightRAG | — | — | — |
| HyperGraphRAG | — | — | — |
| HyperRAG | — | — | — |
| HippoRAG 2 | 65.13 | 90.35 | 94.35 |
| SAG | 80.36 | 93.34 | 96.50 |
03 / ABLATION STUDY
The design choices that help find more evidence
Change one step at a time under the same test conditions and measure how much evidence is lost. These comparisons are called ablations.
MuSiQue · Recall@5
Full SAG80.36%
Triple indexing77.61%
No expansion69.41%
Pointwise reranker67.11%
Follow connections to recover missing facts
Searching for the same person or project adds facts missed by the first retrieval. This raises Recall@5 from 69.41% to 80.36%.
Judge the passages together
Each passage may supply only part of an answer. Considering them together helps select the combination needed, rather than scoring each passage alone.
04 / ARCHITECTURE
Keep the full event and its connections
01 / DATA STRUCTURE
Two records from the same passage
Source passageKeep the text and its source
EventMeaning, relations and conditions
EntitiesPeople, companies and projects
One event linked to several entities
Events preserve meaning; entities support indexing and expansion. Both come from the same passage and are then linked.
One person can connect several documents
An event linked to several entities forms a hyperedge. At query time, shared names connect it to other events relevant to the question.
Find connections at query time without a prebuilt global knowledge graphKeep adding new documents
Process new documents passage by passage and append their records. Combine direct retrieval with connected events, then return the original sources.
Process additions passage by passage and retain source references05 / DEPLOYMENT
Deployment and use
Preparation
Use models to summarize events, extract entities and generate embeddings. Process new documents and append their records.
Use cases
Knowledge bases and agents that need facts across documents. Retrieve connected events, then return their original sources.
Limitations
Manage entity aliases and update stale events separately. The current implementation does not do this automatically.
06 / PERSONAL APPLICATION
SAG
Turn personal files into a local knowledge base for search, cited answers and agent access. Check an answer against its cited sources.
- 01
Add personal files
Organize material in a local knowledge base
- 02
Search or ask
Find relevant content, also accessible to agents
- 03
Check the source
Follow result links and answer citations to the text