WEBVTT

1
00:00:00.350 --> 00:00:01.409
Meet SpectRAG.

2
00:00:01.409 --> 00:00:05.114
Stop retrieving chunks, and start retrieving context.

3
00:00:06.014 --> 00:00:07.622
Ask a multi-hop question,

4
00:00:07.622 --> 00:00:12.044
and classic top-k retrieval grabs six near-duplicate chunks from one cluster.

5
00:00:12.044 --> 00:00:16.063
It finds only one of the three pieces of evidence.

6
00:00:16.063 --> 00:00:18.475
SpectRAG seeds, expands over the graph,

7
00:00:18.475 --> 00:00:21.690
diffuses relevance, and selects a diverse, connected set.

8
00:00:21.690 --> 00:00:23.298
Three out of three.

9
00:00:24.198 --> 00:00:27.605
Offline, every document is split into chunks.

10
00:00:27.605 --> 00:00:31.497
Then the chunks are connected by six signals:

11
00:00:31.497 --> 00:00:33.444
adjacency, sections, semantic similarity,

12
00:00:33.444 --> 00:00:36.363
shared entities, utility questions, and citations.

13
00:00:36.363 --> 00:00:39.770
Together, they form one weighted chunk graph.

14
00:00:40.670 --> 00:00:43.093
Now every chunk has two geometries.

15
00:00:43.093 --> 00:00:45.920
Its embedding captures what the chunk says.

16
00:00:45.920 --> 00:00:47.132
Its spectral coordinates,

17
00:00:47.132 --> 00:00:50.362
from the graph Laplacian, capture where it lives.

18
00:00:50.362 --> 00:00:52.786
Communities pull together, whatever their wording.

19
00:00:54.170 --> 00:00:56.826
At query time, relevance diffuses along the graph.

20
00:00:56.826 --> 00:01:00.147
Bridge chunks that never matched the query still light up.

21
00:01:00.147 --> 00:01:02.804
That is how SpectRAG finds the missing link.

22
00:01:06.170 --> 00:01:08.187
Instead of a fixed top-k,

23
00:01:08.187 --> 00:01:10.205
SpectRAG fills a twelve-thousand-token budget.

24
00:01:10.205 --> 00:01:14.644
A protected core comes first, then coverage-aware picks with no redundancy,

25
00:01:14.644 --> 00:01:18.276
and finally a backfill, so no space is wasted.

26
00:01:20.170 --> 00:01:21.136
The result:

27
00:01:21.136 --> 00:01:25.966
91.8 percent gold-evidence recall, 95 percent accuracy on inference questions,

28
00:01:25.966 --> 00:01:29.830
and an index six times smaller than LightRAG.

29
00:01:30.730 --> 00:01:31.097
SpectRAG.

30
00:01:31.097 --> 00:01:34.776
Retrieve a context set, not a pile of similar chunks.

31
00:01:34.776 --> 00:01:37.350
Read the paper, and try the code.