Construction of shared paper database

The original files of the shared paper database are stored in documents/papers. The first-level subdirectories of this paper repository are the user_id of the lab members.

Labridge utilizes the Parsed content and information to construct a comprehensive and detailed shared literature database, supporting various retrieval methods.

Paper content vector database

Labridge constructs a vector database (VectorIndex) for all shared papers and records the metadata of each paper, as well as the owner of the paper.

Refer to Code docs Func_modules.paper.store.shared_paper_store for details of the construction of the content vector database.

Summarize

Labridge uses LLM to summarize the MainText and Methods sections of each paper added to the shared literature database and constructs a corresponding SummaryVectorIndex. The summary of the MainText focuses on the overall content of the article, main innovations, etc., while the summary of the Methods focuses on the technical approaches used in the article.

• Refer to Code docs Func_modules.paper.synthesizer.summarize for details about content summarizing.
• Refer to Code docs Func_modules.paper.store.paper_store for details about the constructed SummaryVectorIndex.
• The prompts for summarizing MainText and Methods is shown in Code func_modules.paper.prompt.synthesize.paper_summarize

Summarize the paper directories

Labridge uses LLM to recursively generate summaries for each level of the paper repository’s directory, such as the research fields involved in the papers under each directory. This information serves as an important reference for Labridge to recommend papers to lab members and to insert new papers into the shared paper repository.

For details about summarizing the paper directories, refer to Code docs Func_modules.paper.store.paper_store

Shared paper note vector database

The shared paper note vector database is structured as follows. It is structured as a flatten tree, nodes in the first layer record the paper DOI. Each DOI node represents a unique paper. The child nodes of a DOI node record non-overlapped paper content chunks with short length, for the sake of a fine-grained retrieval. The bottom nodes record the notes of users.