A zero-shot combined image retrieval method based on thought chain reasoning of user intent

Abstract

The application belongs to the technical field of machine learning, computer vision and image retrieval, and discloses a zero-shot combined image retrieval method based on thought chain reasoning of user intention, which is realized by using a training-free framework as a whole. The reasoning process based on the thought chain can guide a multi-modal large language model to deeply understand the user retrieval intention, fully excavate key information conducive to retrieval, and effectively exert the reasoning capability of the multi-modal large model without introducing additional training cost. Meanwhile, the multi-source information fusion retrieval method proposed by the application can efficiently integrate various heterogeneous information, reduce visual information loss, and alleviate the negative impact of the large model illusion, thereby significantly improving the retrieval performance. Overall, the application realizes accurate modeling and efficient retrieval of the user retrieval intention through multi-stage reasoning and multi-source information fusion, and has good practical application value and market prospect.

Description

Technical Field

[0001] This invention relates to the fields of machine learning, computer vision, and image retrieval technology, and in particular to a zero-sample combined image retrieval method based on thinking chain inference of user intent. Background Technology

[0002] With the rapid development of multimedia technology and visual-language models, users are demanding greater flexibility and expressiveness from image retrieval systems in practical applications. Traditional image retrieval tasks typically rely on matching based on single-modal information, making it difficult to accurately depict the complex and diverse search intentions of users. To address this, researchers have proposed a combined image retrieval technique. This technique utilizes both reference images and modified text to construct a multimodal query, retrieving target images from image databases that match the user's comprehensive intent. This technology has broad application prospects in scenarios such as multimedia analysis, virtual try-on, product recommendation, and smart retail.

[0003] However, combined image retrieval technology still faces significant challenges in practical applications. Existing methods typically rely on manually annotated triplet data for model training, i.e., labeled samples consisting of a reference image, modified text, and the target image. Constructing such data is costly and struggles to cover complex and ever-changing user needs; data scarcity has become a key bottleneck restricting the development of combined image retrieval technology.

[0004] To alleviate reliance on manually labeled triplet data, researchers have proposed zero-shot combined image retrieval techniques. These methods typically utilize pre-trained visual-language models or large language models to achieve combined image retrieval without specific training or relying solely on image-text pairs. Existing zero-shot combined image retrieval methods primarily transform the combined retrieval problem into a text-to-image retrieval problem. Then, they leverage the cross-modal retrieval capabilities of pre-trained models for the retrieval process.

[0005] Specifically, early zero-shot combined image retrieval methods typically introduced image inversion networks to map reference images into pseudo-word representations, which were then concatenated with modified text to form the query text. However, these methods require a large amount of image-text pairs during the training phase, and the simple concatenation method lacks the ability to explicitly infer user intent, resulting in limited retrieval performance in complex scenarios.

[0006] With the improvement of the reasoning capabilities of large language models, a new class of training-free zero-shot combined image retrieval methods based on visual understanding and language reasoning has emerged in recent years. These methods typically employ a two-stage processing flow: first, a text description is generated for the reference image; then, this description and the modified text are input into a large language model for reasoning to generate the target description or query text for retrieval. While this type of method improves retrieval performance to some extent, it still has several shortcomings: (1) In the process of expressing visual information in text form, the fine-grained visual information of the reference image is inevitably lost. Moreover, image descriptions generated solely based on the reference image usually do not fully integrate the modified text, which is prone to deviating from the user's focus and thus affecting the quality of subsequent reasoning. (2) Existing methods still make relatively limited use of large language models, mostly only used to generate target image descriptions, failing to fully explore their potential in assisting information reasoning. (3) Existing methods often rely excessively on the reasoning results of large language models in the retrieval stage, while ignoring the effective information contained in the original reference image and modified text, resulting in the system being highly sensitive to the output of the language model, easily affected by illusion problems, and limited overall performance and stability.

[0007] In summary, existing zero-shot combined image retrieval techniques still have shortcomings in terms of visual information preservation, user intent inference, and utilization of multi-source information. How to fully utilize the inference capabilities of multimodal large models, reduce visual information loss, and effectively integrate multi-source information from reference images, modified text, and inference results without increasing additional training costs, thereby achieving efficient and robust zero-shot combined image retrieval, remains a pressing technical problem to be solved in this field. Summary of the Invention

[0008] To address the shortcomings of existing methods, this invention aims to propose a zero-shot combined image retrieval method based on thought chain reasoning of user intent. This method is training-free, fully understands user intent, and efficiently integrates multi-source information.

[0009] This method employs a training-free framework, primarily comprising two processes: reasoning based on thought chains and retrieval based on multi-source information fusion. Through multi-stage reasoning and information fusion, it achieves accurate characterization of user search intent and efficient retrieval. It involves multimodal retrieval, multimodal large language models, thought chain reasoning, multi-source information fusion retrieval strategies, and training-free methods; specifically, it is a zero-shot combined image retrieval method based on thought chain reasoning to infer user intent.

[0010] The technical solution of this invention: A zero-shot combined image retrieval method based on user intent reasoning using thought chain includes reasoning based on thought chain and fusion retrieval based on multi-source information. The specific steps are as follows: Step 1: Encode and embed the candidate images according to the CLIP image encoder; Step 2: Perform reasoning based on thought chains; The reasoning process is divided into 6 sub-tasks, including understanding the reference image, analyzing and modifying the text, classifying visual elements, generating relevant region descriptions, executing modification instructions, and generating target image descriptions. After completing the 6 sub-tasks, 4 pieces of information are obtained: target image description, relevant region description, positive visual elements, and negative visual elements. Step 3: Process multi-source information; Encode and embed the reference image according to the CLIP image encoder, and encode the modified text, target image description, relevant region description, positive visual elements, and negative visual elements according to the CLIP text encoder; Calculate the similarity between each item and the candidate image. Step 4: Input the similarity calculated from the reference image, modified text, target image description, and relevant region description into the global matching module to obtain the global search results; Step 5: Input the similarity calculated from the positive and negative visual elements into the local enhancement module to obtain the local retrieval results; Step 6: Weighted fusion of global and local search results.

[0011] The reasoning based on the thought chain is implemented through a multimodal large language model; the reference image and modified text are input into the multimodal large language model, and six reasoning sub-tasks are performed through prompt word templates; The understanding of the reference image specifically involves: comprehensively understanding the reference image, identifying and listing all visible visual elements in the reference image, including specific objects, background information, and shooting perspective; the analysis must be strictly based on the reference image itself and must not introduce any content or inferred information that does not actually appear in the reference image. The analysis and modification of the text specifically involves: systematically analyzing the modified text to identify which visual elements were modified and the specific modification methods; classifying all modifications into absolute modifications and relative modifications; absolute modifications refer to modifications that are clear and specific in themselves and can be understood without comparison with the original state; relative modifications, on the other hand, require comparison with the original state to clarify the modification results; checking whether changes in background, perspective, or number of objects are involved; when the modified text describes a change in the entire scene, it is considered a modification step, and all elements in the reference image are considered to have changed; the modified text is also broken down into multiple modification steps based on the modified objects; The classification of visual elements specifically involves dividing all visual elements in the reference image into relevant elements and irrelevant elements. Relevant elements are those that may be affected by the modification or those that are explicitly required to be retained during the modification. Irrelevant elements are the remaining elements in the reference image that do not meet the above conditions. When the modified text involves changes to the entire scene, all visual elements in the reference image are considered relevant elements. The specific method for generating a relevant region description is as follows: based on the identified relevant elements, generate a short and clear image description that focuses only on the relevant elements themselves, maintains the simplicity of expression, and avoids redundant or irrelevant information. The specific steps for executing the modification instructions are as follows: Apply the modifications step-by-step to the description of the relevant areas; during this process, clearly explain the understanding of the modification intent, and maintain the original context and overall coherence of the reference image while meeting the modification requirements; all modifications must be logical, ensuring a reasonable semantic connection before and after the modification; for relative modifications, the final result must include both the original state and the modified state; and the results before and after the modification should be presented in the form of a comparison list, in the following format: [[Before 1, After 1], [Before 2, After 2], …] ; The specific steps for generating the target image description are as follows: based on the complete modification process, generate a coherent and concise target image description; the target image description accurately reflects all modification results and maintains conciseness; it must not mention any content that does not appear in the target image.

[0012] The similarity calculation uses cosine similarity.

[0013] The global matching module specifically comprises: given a global input , This represents the similarity score calculated using a reference image; This represents the similarity score calculated using the modified text; This represents the similarity score calculated using the target image description; This represents the similarity score calculated using relevant region descriptions; Based on their characteristics, they are divided into two groups; the first group consists of... and Composition; the second group includes and The global aggregation result is defined as: in, It is a hyperparameter used to balance the contributions of various factors.

[0014] The local enhancement module specifically involves: spatially coupling positive and negative visual elements in pairs to obtain paired representations; given local input... and ,in It is the number of positive and negative visual element pairs. It is the number of candidate images. , Representatives used the first A similarity score is calculated from each positive visual element. Representatives used the first A similarity score is calculated for each negative visual element; the initial result of the similarity score for each pair of positive and negative visual elements is first calculated using the Softmax function to determine its weight: in This is equivalent to performing binary classification on each candidate image to distinguish between positive and negative visual elements, and then enhancing the original similarity score based on the obtained weights to obtain the enhanced result. An image-to-text matching method is introduced to achieve better local alignment. A skip connection is used to fuse the initial result with the enhanced result. The entire process can be represented as: Finally, a simple weight is adopted. Weighted fusion of global and local search results: ; The value is adjusted based on downstream data.

[0015] Compared with the prior art, the present invention has the following beneficial effects: (1) The reasoning process based on thought chain proposed in this invention can more fully utilize the reasoning ability of multimodal large language models and mine richer and more useful information for retrieval. This enables the model to fully understand user intent and effectively improve the retrieval recall rate.

[0016] (2) The multi-source information-based fusion retrieval process proposed in this invention does not only rely on the original input or the reasoning results of multimodal models, but also makes full use of multi-source information; it does not only rely on global matching, but also considers using local information to improve the retrieval effect. This makes the model improve performance while significantly reducing the impact of the illusion of a large multimodal model, making the retrieval results more stable. Attached Figure Description

[0017] Figure 1 This is a technical framework diagram of the present invention.

[0018] Figure 2 This is the complete prompt for the present invention.

[0019] Figure 3 This is a structural diagram of the global matching module and the local enhancement module of the present invention; (a) is the global matching module, and (b) is the local enhancement module.

[0020] Figure 4 This is a schematic diagram of the thought chain reasoning process in a specific embodiment. Detailed Implementation

[0021] A zero-shot combined image retrieval method based on user intent reasoning using thought chain is proposed. The method consists of two processes: reasoning based on thought chain and fusion retrieval based on multi-source information. The specific steps are as follows: Step 1: This invention is a training-free method, therefore no training data is required. Following the general image retrieval approach, the candidate images are first encoded and embedded using the CLIP image encoder; Step 2: Perform reasoning based on thought chain. The entire reasoning process can be divided into 6 sub-tasks: (1) understanding the reference image; (2) analyzing and modifying the text; (3) classifying visual elements; (4) generating relevant region descriptions; (5) executing modification instructions; and (6) generating a target image description. After completing the 6 reasoning sub-tasks, 4 pieces of information are obtained: target image description, relevant image description, positive visual elements, and negative visual elements. This information will be used in the subsequent retrieval process. Step 3: Process multi-source information. The CLIP image encoder is used to encode and embed the reference image, and the CLIP text encoder is used to encode the modified text, target image description, relevant region description, positive visual elements, and negative visual elements. Then, the similarity between each item and the candidate image is calculated. Step 4: Input the similarity calculated from the reference image, modified text, target image description, and relevant region description into the global matching module to obtain the global search results.

[0022] Step 5: Input the similarity calculated from the positive and negative visual elements into the local enhancement module to obtain the local retrieval results.

[0023] Step 6: Use a simple weighted fusion method to combine the global and local search results.

[0024] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0025] The specific implementation of this invention is as follows: 1. Encoding candidate images After the candidate images are processed using CLIP's own image preprocessing pipeline, they are input into CLIP's image encoder for encoding, and finally the image embedding representations are obtained and stored for later use.

[0026] 2. Reasoning based on thought chains like Figure 1 As shown, part (a) is the reasoning based on thought chains. The reference image and modified text are input into a multimodal large language model. After six reasoning sub-tasks, four pieces of information are obtained, which will be used in the subsequent retrieval process based on multi-source information. These six reasoning sub-tasks will now be described in detail, with corresponding prompts as follows: Figure 2 As shown.

[0027] Task 1: Understanding the Reference Image. This task requires a comprehensive understanding of the reference image, identifying and listing all visible visual elements, including specific objects, background information, and shooting perspective. Furthermore, the analysis must be strictly based on the reference image itself, and must not introduce any content or inferred information that does not actually appear in the image.

[0028] Task 2: Analyze Modified Text. This task requires a systematic analysis of modified text to identify which visual elements were modified and the specific methods of modification. All modifications must be categorized as absolute or relative: absolute modifications are those whose results are clear and specific, understandable without comparison to the original state; relative modifications, on the other hand, require comparison with the original state (e.g., longer, shorter, more, less, different, etc.) to clarify the result. Furthermore, it's necessary to check for changes in background, perspective, or the number of objects; if the modified text describes a change in the entire scene, it's considered a single modification step, and all elements in the reference image are considered to have changed. Modified text should also be broken down into multiple modification steps based on the modified objects.

[0029] Task 3: Classify Visual Elements. This task requires classifying all visual elements in a reference image into relevant and irrelevant elements. Relevant elements are those that may be affected by modifications (including changes in attributes, quantity, relationships, background, or viewpoint), or those explicitly required to be retained in the modification; irrelevant elements are the remaining elements in the image that do not meet the above criteria. It is important to note that if the modification involves changes to the entire scene, then all visual elements in the reference image should be considered relevant elements.

[0030] Task 4: Generate Relevant Region Descriptions. This task requires generating a concise and clear image description based on the identified relevant elements, focusing only on the relevant elements themselves and maintaining simplicity as much as possible, avoiding redundant or irrelevant information.

[0031] Task 5: Execute the modification instructions. This task requires applying the modifications step-by-step to the relevant area descriptions. During this process, a clear explanation of the intended modification must be provided, maintaining the original context and overall coherence of the reference image while meeting the modification requirements. All modifications should be logical, ensuring a reasonable semantic connection between the original and modified states; for relative modifications, the final result must include both the original and modified states. Furthermore, a comparison list of the results before and after modification should be provided, in the following format: [[Before 1, After 1], [Before 2, After 2], …] .

[0032] Task 6: Generate a description of the target image. This task requires generating a coherent and concise description of the target image based on the complete modification process. The description must accurately reflect all modification results while maintaining conciseness. Furthermore, it must not mention any content not present in the target image.

[0033] Guided by the above six sub-tasks, the model reasoned step by step and obtained the target image description, related image description, positive visual elements, and negative visual elements, which will be used in the subsequent retrieval process. Furthermore, to further standardize the reasoning process and unleash the reasoning capabilities of the multimodal large language model, this invention also provides a few examples to help the multimodal large language model better understand the reasoning process.

[0034] 3. Processing multi-source information The current information consists of images (reference images) and text (including modified text, target image descriptions, related image descriptions, positive visual elements, and negative visual elements). These are encoded and embedded using CLIP's image encoder and text encoder, respectively.

[0035] Formalistically, the reference image is denoted as The modified text is recorded as The target image description is denoted as The relevant description is recorded as .in, Representing image space, Represents the text space. The set of positive sample elements is defined as follows: The set of negative sample elements is defined as ,in This represents the number of positive or negative visual elements. For candidate images in the database, the set is represented as... , This indicates the number of candidate images.

[0036] Calculate the candidate image relative to and The similarity score is used to obtain as well as ,in .

[0037] Taking text modification as an example, the similarity calculation method is expressed as follows: in, and These represent image and text encoders, respectively. This represents the cosine similarity.

[0038] 4. Global Matching Module like Figure 3 As shown in (a), given a global input Based on their characteristics, they are divided into two groups. The first group consists of... and Composition. In reality, multimodal large language models can produce inference errors, and modified text can be viewed to some extent as a simplified description of the target image. Combining the two helps mitigate the impact of errors and guides the model to focus more on the modified region. The second group includes... and Generally, retrieval based on reference images can introduce noise, especially in modified regions. To address this issue, this invention introduces a relevant region description, which accurately characterizes these regions and filters out false matches. Based on the above design, the global aggregation result is defined as: in, It is a hyperparameter used to balance the contributions of various factors.

[0039] 5. Local Enhancement Module While the global aggregation module is effective overall, it may overlook some valuable local details. Therefore, this invention further proposes a local enhancement module. Specifically, positive and negative visual elements possess local perceptual characteristics in space, and thus this invention couples them in pairs. The resulting paired representations provide richer information for retrieval. For example... Figure 3 As shown in (b), for each pair of positive and negative visual elements, this invention first uses the Softmax function to calculate their weights: in Based on this, the present invention performs binary classification on each candidate image to distinguish between positive and negative visual elements, and uses the obtained weights to enhance the original similarity score. Unlike previous methods that only perform text-to-image matching, the present invention further introduces an image-to-text matching method to achieve better local alignment results. To improve stability while preserving the original information, the present invention also uses skip connections to fuse the initial result with the enhanced result. The entire process can be represented as: 6. Weighted fusion like Figure 1 As shown in section (b), a simple weight is ultimately used. Weighted fusion of global and local search results: . The value can be adjusted based on downstream data.

[0040] Figure 4 A specific implementation example of the thought chain reasoning process of this method in the field of fashion apparel retrieval is given to illustrate the step-by-step reasoning process from input to output.

[0041] In this example, the reference image is a full-body frontal photograph: a woman standing against a pure white background, wearing a black and white snakeskin knee-length dress with black high heels. The dress has a V-neck and a black belt. The modified text is: "Change the dress color to red, make it show more bust, and make it more shiny." That is: change the dress color to red to make it more prominent on the bust and enhance its shine. The outputs at each stage of this method's processing flow are as follows: (1) Understanding the reference image The model first performs structured analysis on the image, identifying all visual elements and their attributes, including: a. Object level: woman, skirt, high heels; b. Background: Pure white background; c. Viewpoint: Full-body frontal view; d. Attribute level: The skirt features a black and white snakeskin pattern, V-neck, short sleeves, knee-length, black belt, and matte texture.

[0042] The output of this stage is a complete list of visual elements that does not introduce any additional information.

[0043] (2) Analyze and modify the text The model decomposes the modified text, identifying all modified objects and modification types: a. Item to be modified: Skirt; b. Modification 1: Change the color to red (absolute modification); c. Modification 2: Emphasize the chest more (relative modification, based on the original V-neck design); d. Modification 3: More glossy (relative modification, based on the original matte material); e. Background, perspective, quantity: No change.

[0044] At the same time, the model breaks down the modifications into three independent steps, forming a structured modification sequence.

[0045] (3) Classification of visual elements Based on the scope of the modification's impact, visual elements are divided into: a. Related element: skirt; b. Irrelevant elements: woman, high heels, background, perspective.

[0046] Since this example does not involve a complete scene replacement, only the skirt will be included in subsequent updates.

[0047] (4) Generate relevant region descriptions The model generates a concise description for the "skirt": "A black and white snake print knee-length dress with a V-neckline and a black belt." This description will be used as the baseline state for subsequent modifications.

[0048] (5) Execute the modification command The model is modified sequentially, and before-and-after comparison information is retained: a. Step 1 (Color Modification): Black and white snakeskin skirt → Red skirt b. Second step (neckline adjustment): V-neck design → more revealing neckline design c. Step 3 (Material Adjustment): Matte Material → Glossy Material At each step, the model ensures semantic continuity and does not introduce elements that are inconsistent with the reference image.

[0049] (6) Generate target image description Based on all the modifications, the final target image description is generated as: "A red, shiny dress with a more revealing neckline."

Claims

1. A zero-shot combined image retrieval method based on thought chain reasoning of user intent, characterized in that, This includes reasoning based on thought chains and retrieval based on the fusion of multi-source information. The specific steps are as follows: Step 1: Encode and embed the candidate images according to the CLIP image encoder; Step 2: Perform reasoning based on thought chains; The reasoning process is divided into 6 sub-tasks, including understanding the reference image, analyzing and modifying the text, classifying visual elements, generating relevant region descriptions, executing modification instructions, and generating target image descriptions. After completing the 6 sub-tasks, 4 pieces of information are obtained: target image description, relevant region description, positive visual elements, and negative visual elements. Step 3: Process multi-source information; Encode and embed the reference image according to the CLIP image encoder, and encode the modified text, target image description, relevant region description, positive visual elements, and negative visual elements according to the CLIP text encoder; Calculate the similarity between each item and the candidate image. Step 4: Input the similarity calculated from the reference image, modified text, target image description, and relevant region description into the global matching module to obtain the global search results; Step 5: Input the similarity calculated from the positive and negative visual elements into the local enhancement module to obtain the local retrieval results; Step 6: Weighted fusion of global and local search results.

2. The zero-shot compositional image retrieval method based on reasoning user intent of thought chain according to claim 1, characterized in that, The reasoning based on the thought chain is implemented through a multimodal large language model; the reference image and modified text are input into the multimodal large language model, and six reasoning sub-tasks are performed sequentially through prompt word templates; The understanding of the reference image specifically involves: comprehensively understanding the reference image, identifying and listing all visible visual elements in the reference image, including specific objects, background information, and shooting perspective; the analysis must be strictly based on the reference image itself and must not introduce any content or inferred information that does not actually appear in the reference image. The analysis and modification of the text specifically involves: systematically analyzing the modified text to identify which visual elements were modified and the specific modification methods; classifying all modifications into absolute modifications and relative modifications; absolute modifications refer to modifications that are clear and specific in themselves and can be understood without comparison with the original state; relative modifications, on the other hand, require comparison with the original state to clarify the modification results; checking whether changes in background, perspective, or number of objects are involved; when the modified text describes a change in the entire scene, it is considered a modification step, and all elements in the reference image are considered to have changed; the modified text is also broken down into multiple modification steps based on the modified objects; The classification of visual elements specifically involves dividing all visual elements in the reference image into relevant elements and irrelevant elements. Relevant elements are those that may be affected by the modification or those that are explicitly required to be retained during the modification. Irrelevant elements are the remaining elements in the reference image that do not meet the above conditions. When the modified text involves changes to the entire scene, all visual elements in the reference image are considered relevant elements. The generation of relevant region description specifically involves: generating a concise and clear image description based on the identified relevant elements, focusing only on the relevant elements themselves, maintaining the simplicity of expression, and avoiding redundant or irrelevant information; The specific steps for executing the modification instructions are as follows: Apply the modifications step-by-step to the description of the relevant areas; during this process, clearly explain the understanding of the modification intent, and maintain the original context and overall coherence of the reference image while meeting the modification requirements; all modifications must be logical, ensuring a reasonable semantic connection before and after the modification; for relative modifications, the final result must include both the original state and the modified state; and the results before and after the modification should be presented in the form of a comparison list, in the following format: [[Before 1, After 1], [Before 2, After 2], …] ; The specific steps for generating the target image description are as follows: based on the complete modification process, generate a coherent and concise target image description; the target image description accurately reflects all modification results and maintains conciseness; it must not mention any content that does not appear in the target image.

3. The zero-sample combined image retrieval method based on user intent reasoning using thought chain as described in claim 2, characterized in that, The similarity calculation uses cosine similarity.

4. The zero-sample combined image retrieval method based on user intent reasoning using thought chain as described in claim 3, characterized in that, The global matching module specifically comprises: given a global input , This represents the similarity score calculated using a reference image; This represents the similarity score calculated using the modified text; This represents the similarity score calculated using the target image description; This represents the similarity score calculated using relevant region descriptions; Based on their characteristics, they are divided into two groups; the first group consists of... and Composition; the second group includes and The global aggregation result is defined as: in, It is a hyperparameter used to balance the contributions of various factors.

5. The zero-sample combined image retrieval method based on user intent reasoning using thought chain inference as described in claim 4, characterized in that, The local enhancement module specifically involves: spatially coupling positive and negative visual elements in pairs to obtain paired representations; given local input... and ,in It is the number of positive and negative visual element pairs. It is the number of candidate images. , Representatives used the first A similarity score is calculated from each positive visual element. Representatives used the first A similarity score is calculated for each negative visual element; the initial result of the similarity score for each pair of positive and negative visual elements is first calculated using the Softmax function to determine its weight: in This is equivalent to performing binary classification on each candidate image to distinguish between positive and negative visual elements, and then enhancing the original similarity score based on the obtained weights to obtain the enhanced result. An image-to-text matching method is introduced to achieve better local alignment. A skip connection is used to fuse the initial result with the enhanced result. The entire process can be represented as: Finally, a simple weight is adopted. Weighted fusion of global and local search results: ; The value is adjusted based on downstream data.

Images