PASSAGE A · PRODUCTPENDING
Nora support follows policy R-17.
SAG2
Lighter to deploy. Faster to index. Better at finding evidence.
Index the text, then retrieve the facts you need
SAG2 builds on SAG's approach to connecting documents. It indexes the original text by keywords, removing LLM processing and embeddings from preparation. When a question arrives, it finds useful passages and follows their terms to retrieve more evidence.
- Lighter: one database file
- Keep text and indexes in a SQLite file, without a separate database server or vector store.
- Faster: no models needed for indexing
- Build a keyword index directly from the text. In the paper's experiment, indexing 11,656 passages took 2.36 seconds, with no LLM or embedding calls.
- More accurate: connect source passages
- Use what has already been found to retrieve more of the facts a complex answer needs. Compare results on three public benchmarks below.
86.62%Evidence found in the top tenMuSiQue · Recall@10
81.10%Enterprise document recall511,962 documents · Recall@10
2.36 sTime to index 11,656 passagesMuSiQue · No indexing model calls
Results from the papers. Benchmarks, metrics and test conditions are explained below.
01 / HOW IT WORKS
Take a term from a passage and search again
SAG2 indexes the original text. At query time, BM25 retrieves passages, JEV selects useful evidence, and terms from that evidence drive another search.
SEARCH–JUDGE–EXPAND1 / 5
How long is Nora's warranty?
PREPARATION
Source text + keyword index → one SQLite file
Local text indexing, with no model callsBM25JEVTERMS
New terms return to BM25 for another search
READY TO SEARCHNoraFrom the original question
PASSAGE B · DESIGNPENDING
Nora packaging comes in two colors.
Related passages have not been searched yet
CONTEXT FOR JEV
Question + selected passage A → inform the next judgment
Evidence recovered: Nora → R-17 → two years
Index the original text
Store source passages and term occurrences in SQLite. Build the full-text index locally, with no model calls.
A fictional example illustrating the retrieval process. Select any step to inspect it; autoplay stops after one pass.
VIDEO / SAG2
SAG2 explained
See how embedding-free indexing, BM25, JEV and successive searches work together.
Video and resources on GitHub02 / BENCHMARKS
Find more supporting facts on three public benchmarks
Recall@5 measures the required evidence found in the first five results; Recall@10 looks at the first ten. Recovering more of those facts gives an agent a more complete basis for its answer.
MuSiQueRecall@10 · Higher is better
+3.25 points vs SAG
SAG286.62%
SAG83.37%
NV-Embed-v277.27%
GritLM-7B75.19%
HippoRAG 273.76%
GTE-Qwen2-7B70.80%
BGE-Large-EN-v1.564.08%
BM2554.41%
Contriever53.58%
1,000 fixed development questions per benchmark; MuSiQue was also used during development. Values follow the SAG2 paper. SAG and SAG2 use different models and implementations, so these compare the complete systems.
Full benchmark dataRecall@10
| Method | MuSiQue | 2WikiMultiHopQA | HotpotQA |
|---|---|---|---|
| BM25 | 54.41 | 71.50 | 86.90 |
| Contriever | 53.58 | 65.20 | 82.75 |
| BGE-Large-EN-v1.5 | 64.08 | 73.75 | 93.35 |
| GTE-Qwen2-7B | 70.80 | 78.63 | 93.55 |
| GritLM-7B | 75.19 | 79.40 | 97.20 |
| NV-Embed-v2 | 77.27 | 80.75 | 97.70 |
| HippoRAG 2 | 73.76 | 93.40 | 97.15 |
| SAG | 83.37 | 93.57 | 97.70 |
| SAG2 | 86.62 | 97.43 | 99.00 |
03 / ABLATION STUDY
The design choices that help find more evidence
Use the same 1,000 MuSiQue questions to test fewer search rounds and judgments made without the passages already found.
MuSiQue · Recall@10
Full SAG2 · two expansions86.62%
One expansion83.36%
No shared context76.25%
No expansion · BM25 + JEV67.93%
One more search round finds more evidence
Initial retrieval reaches 67.93% Recall@10. One additional search round raises it to 83.36%, and two reach 86.62%.
Use known facts to assess new passages
Judging new passages without the facts already found lowers Recall@10 from 86.62% to 76.25%.
04 / SCALE & PREPARATION
Fast to index, tested on half a million documents
Preparation builds a keyword index directly from the text, without LLMs or embedding models. Below are the indexing measurements and retrieval results on more than half a million documents.
01 / INDEXING COST
Measured indexing cost on MuSiQue
2.36s
To index 11,656 source passages by keyword
- Source passages
- 11,656
- Database file
- 25.87 MiB
- Indexing model calls
- 0
Timing comes from the paper's experiment; it varies with hardware and the documents being indexed.
02 / RETRIEVAL AT SCALE
EnterpriseRAG-Bench
81.10%
Required documents found in the top ten · Recall@10
BM25 68.40%Vector 46.00%
511,962 synthetic documents model company emails, messages, meetings and other records. 470 questions identify the documents needed. Baselines come from the dataset paper with different compute budgets; this measures document retrieval, not generated answers.
05 / DEPLOYMENT
Deployment and use
Local deployment
Keep source text and keyword indexes in one SQLite file. Indexing needs no LLM or embedding model calls.
Retrieval and answers
JEV judges passages during semantic retrieval; useful passages drive further searches. Connect your own model to generate answers.
Validate before use
The paper uses fixed test sets. Measure retrieval quality and online speed on your own documents and hardware.
06 / RELATED PRODUCT
SAG Wiki
A local knowledge base powered by SAG2. Choose a folder to search supported documents, notes and code, then ask questions and check the cited source passages.
- 01
Choose a folder
Add supported documents, notes and code
- 02
Search or ask
Retrieve the relevant passages with SAG2
- 03
Check citations
Return from the answer to the original passages