Personal experiment log retrieval

The experiment log is related to time, therefore, Labridge adopts a multi-level retrieval + timestamp filtering method based on the storage structure of the experiment log.

The first retrieval step

In the first retrieval step, Labridge retrieves the most likely experiment_top_k experiments based on the similarity between the Query vector and the descriptions of all experiment nodes. At the same time, Labridge searches within all nodes of log types to retrieve the most similar first_top_k log nodes based on similarity, and obtains the corresponding experiment nodes for them.

The retrieved experiment nodes will serve as the scope for the next retrieval step.

The second retrieval step

Within the range of experiment nodes obtained from the first retrieval step, Labridge retrieves second_top_k log nodes from the log nodes of these experiment nodes to serve as the retrieval results.

Timestamp filtering

In the second retrieval step, Labridge will perform timestamp filtering based on the input start and end times (if provided). The ultimately retrieved experiment log content will be provided to LLM as input.

Refer to Code docs Func_modules.memory.experiment.retrieve_log for details of retrieval