# US 2026/0099554 A1 · Method and apparatus for searching multi-vector using span-level sequence compression technique

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2026/0099554 A1
- Document ID: US20260099554A1
- Status: Published application
- Kind: A1
- Application number: 19/415,303
- Filed: 2025-12-10
- Published: 2026-04-09
- Earliest claimed: 2023-10-18
- Assignees: SK Telecom Co., Ltd.
- Applicants: SK Telecom Co., Ltd.
- Inventors: Cheon Eum PARK; Kyung Lang PARK; Soon Woong LEE; Yong Hun LEE; Seung Min HEO
- Web page: https://patentgrep.com/patent/US20260099554A1

## CPC

- G06F 16/93: Document management systems
- G06F 16/33: Querying
- G06F 16/332: Query formulation
- G06F 16/3347: {using vector based model}
- G06F 17/16: Matrix or vector computation {, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization (matrix transposition G06F7/78)}
- G06F 40/284: Lexical analysis, e.g. tokenisation or collocates
- G06N 3/0455: Auto-encoder networks; Encoder-decoder networks
- G06N 3/08: Learning methods
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

An apparatus according to an embodiment of the present disclosure comprises: at least one memory for storing instructions; and at least one processor. The at least one processor is configured to execute the instructions to perform the steps of: generating a span including a span index for an arbitrary document token; encoding an input query token to generate a query token encoding vector; encoding a document token; generating a document token encoding vector; compressing the document token encoding vector into a span vector according to the span index and outputting the compressed document token encoding vector; and searching for a document token encoding vector having the highest similarity to the query token encoding vector by calculating a score between the query token encoding vector and the compressed document token encoding vector.

## Claims

### Claim 1

An apparatus for searching multi-vector using a span-level sequence compression technique, comprising:
at least one memory storing instructions; and
at least one processor,
wherein the at least one processor configured to execute the instructions to:
generate a span comprising a span index for any document token;
encode an input query token to generate a query token encoding vector;
encode the document token to generate a document token encoding vector;
compress the document token encoding vector into a span vector according to the span index to output a compressed document token encoding vector; and
calculate a score between the query token encoding vector and the compressed document token encoding vector and retrieve a document token encoding vector having the highest similarity to the query token encoding vector.

### Claim 2 (depends on 1)

The apparatus according to claim 1, wherein
the at least one processor is further configured to:
calculate a first loss value using cross-entropy on the score;
calculate a document score corresponding to a single vector;
calculate a second loss value using cross-entropy on the document score corresponding to the single vector; and
perform training by adding the first loss value and the second loss value.

### Claim 3 (depends on 2)

The apparatus according to claim 2, wherein
the document score corresponding to the single vector is calculated by:
extracting, from the query token encoding vector and the document token encoding vector, vectors at a first token position; and
performing a dot-product on the extracted query token encoding vector and the extracted document token encoding vector.

### Claim 4 (depends on 1)

The apparatus according to claim 1, wherein
the span is generated by performing window sliding according to span width and sliding step rate, and
the span index is generated based on a total number of the spans.

### Claim 5 (depends on 1)

The apparatus according to claim 1, wherein
the number of the spans is configured to be equal to a number of tensors used when indexing the input document tokens.

### Claim 6 (depends on 1)

The apparatus according to claim 1, wherein
the span vector comprises:
a vector obtained by calculating the alignment scores of the start token, the end token, and the tokens included within the span range among the tokens corresponding to the span range, and concatenating the calculated alignment scores.

### Claim 7 (depends on 1)

The apparatus according to claim 1, wherein
the score is calculated using a MaxSim function, and
the MaxSim function satisfies the following mathematical equation:
f
⁡
(
Q
,
D
s
)
=
∑
i
=
1
n
∑
k
=
1
l
A
ik
⁢
q
i
⁢
s
k
[
Mathematical
⁢
equation
]
where qi represents the query token encoding vector, sk represents the span vector, and Aik represents the maximum value masking vector.

### Claim 8 (depends on 7)

The apparatus according to claim 7, wherein
the calculating the score comprises:
calculating a score corresponding to a query-document pair when, in the MaxSim function, Aik is 1 and qi, sk has a maximum value.

### Claim 9

A computer-implemented method comprising:
generating a span comprising a span index for any document token;
encoding an input query token to generate a query token encoding vector;
encoding the document token to generate a document token encoding vector;
compressing the document token encoding vector into a span vector according to the span index to output a compressed document token encoding vector; and
calculating a score between the query token encoding vector and the compressed document token encoding vector and retrieving a document token encoding vector having the highest similarity to the query token encoding vector.

### Claim 10 (depends on 9)

The computer-implemented method according to claim 9, further comprising:
calculating a first loss value using cross-entropy on the score;
calculating a document score corresponding to a single vector;
calculating a second loss value using cross-entropy on the document score corresponding to the single vector; and
performing training by adding the first loss value and the second loss value.

### Claim 11 (depends on 10)

The computer-implemented method according to claim 10, wherein
the document score corresponding to the single vector is calculated by:
extracting, from the query token encoding vector and the document token encoding vector, vectors at a first token position; and
performing a dot-product on the extracted query token encoding vector and the extracted document token encoding vector.

### Claim 12 (depends on 9)

The computer-implemented method according to claim 9, wherein
the span is generated by performing window sliding according to span width and sliding step rate, and
the span index is generated based on a total number of the spans.

### Claim 13 (depends on 9)

The computer-implemented method according to claim 9, wherein
the number of the spans is configured to be equal to a number of tensors used when indexing the input document tokens.

### Claim 14 (depends on 9)

The computer-implemented method according to claim 9, wherein
the span vector comprises:
a vector obtained by calculating the alignment scores of the start token, the end token, and the tokens included within the span range among the tokens corresponding to the span range, and concatenating the calculated alignment scores.

### Claim 15 (depends on 9)

The computer-implemented method according to claim 9, wherein
the score is calculated using a MaxSim function, and
the MaxSim function satisfies the following mathematical equation:
f
⁡
(
Q
,
D
s
)
=
∑
i
=
1
n
∑
k
=
1
l
A
ik
⁢
q
i
⁢
s
k
[
Mathematical
⁢
equation
]
where qi represents the query token encoding vector, sk represents the span vector, and Aik represents the maximum value masking vector.

### Claim 16 (depends on 15)

The computer-implemented method according to claim 15, wherein
the calculating the score comprises:
calculating a score corresponding to a query-document pair when, in the MaxSim function, Aik is 1 and qi, sk has a maximum value.

### Claim 17

A non-transitory computer-readable recording medium having instructions stored thereon, wherein the instructions, when executed by a computer, cause the computer to:
generate a span comprising a span index for any document token;
encode an input query token to generate a query token encoding vector;
encode the document token to generate a document token encoding vector;
compress the document token encoding vector into a span vector according to the span index to output a compressed document token encoding vector; and
calculate a score between the query token encoding vector and the compressed document token encoding vector and retrieve a document token encoding vector having the highest similarity to the query token encoding vector.