Deep Research

Note: Deep Research is a Benchling AI feature. See the AI at Benchling page for information on how to enable Deep Research and other AI-based features.

This feature is in Preview, so keep in mind that it is a work in progress. Usage limits may apply as we prepare for general availability.

Overview

Deep Research helps answer complex questions by retrieving and analyzing data across your Benchling tenant. It can extract information from unstructured text (whether in Notebook Entries or Attachments) and combine it with structured data (Entities, Results, and more) to produce comprehensive, citation-backed reports.

Getting started

Deep Research is located in the AI section of our sidebar.

1. Navigate to the AI icon on the sidebar
2. Click on the Deep Research toggle
3. Enter your prompt into the input box at the bottom
4. (Optionally) attach any files and include public data 
5. Click Send ⬆️

Responses can take anywhere from a couple of minutes up to half an hour depending on query complexity and data volume. If you don't need as comprehensive an answer and want a faster response, check out Ask.

Understanding Deep Research Output

Deep Research reports contain two types of references:

1. Benchling Object References: Appear as interactive chips (similar to mentions in Notebook Entries)
   • Click to navigate directly to the source

   • Hover to preview key information

     

      

2. Data Citations: Show up as bracketed end-note style superscript links [#] that link to specific actions Deep Research took to gather information
   • Click citations to view source context

You can use both types of references to validate the correctness of the Deep Research report.

Deep Research reports can be saved as .docx Word files, appended to the end of an existing Notebook Entry, or as a new Entry or Template by clicking the Save as button at the bottom of the report.

Reviewing and Exporting Citations and Plots

Deep Research's core tool set is similar to and interacts with the same tools you have access to in the Benchling UI. The Tool Details modal will show you the input and output of each step in Deep Research's process. If you would like to review the same data in the Benchling UI yourself you can do so by following these steps for the corresponding type of tool call:

• Searching Benchling, Searching Notebook Entries, Reading Notebook, and Looking things up
  • These are all either search queries and their results or reading the contents of a Benchling Notebook Entry, its attachments, a Benchling Entity (sequence, custom entity) or other Benchling object (task).
    • Search Queries will be formatted like Query: "CL001,CL002" 
      • Replicate this by pasting each item in the comma separated list into Benchling search. 
    • Reading Notebook or Looking things up will typically return the Entry Title, EXP ID or both near the top of the input. If you see an API ID like etr_... bfi_... or seq_... you can use global search > Filters > "API ID" and then search for one or more (comma separated) API IDs 
• Reading Attachment
  • These refer to files attached to your Deep Research chat. Files attached to Benchling Notebook entries will be referenced in "Reading Notebook" citations. 
• Querying Warehouse
  • These should be formatted like Query: "SELECT ..."
  • Create a new block in dashboard in the Insights application and replace the default query body with the query between the double quotation marks and then click "Run query." 
  • You can use SQL Writer to modify this query for other, similar use cases. 
• "Creating chart" for custom plots, charts, or graphs
  • The Input will contain the python code for a function named "llm_make_figure"
    • This will include the data that will be incorporated in the graphic(s) itself, not necessarily the raw data e.g. if the average values are being graphed, the code likely only contains the averages, not the datapoints from which they were derived. 
    • You can insert this function into any python context and perform actions on the figure including exporting to several image formats. 
    • For detailed instructions, see How To Export Custom Plots from Deep Research Citations. 

Accessing Previous Research

Any Deep Research query you have made can be returned to at a later time by following these steps:

1. Navigate to the AI icon on the sidebar
2. Click on AI Chats at the top of the sidebar
3. Click on a chat to open it
4. Ask follow-up questions to refine or expand previous research

Deep Research with Public Data

When using Deep Research with public data search, you can ask questions across both your Benchling data and public sources. Deep Research automatically crafts a query using your data, which is then used to query the internet. 

Deep Research with public data search is disabled by default. To enable, go to your AI settings in the tenant admin console. Once enabled, start a Deep Research chat and toggle the option to include public data.
 

Prompt Engineering Guide

Deep Research is very flexible and can answer a wide array of questions no matter what you input. However, there are a few ways you can get better output from Deep Research.

Here's an example prompt structure that can help:

[Goal] + [Specific Data/Entities] + [Output Format]

You should also keep the following in mind as you are constructing your prompts:

• Set Acceptance Criteria: Include specific criteria or thresholds for success
• Be Specific: Reference exact entity IDs, experiment names, or date ranges
• Request Format: Specify if you want tables, links, or narrative summaries
• Iterate: Start broad, then refine with follow-up questions

Common Use Cases with Example Prompts 

Here are a few use cases and example prompts we've seen Deep Research be helpful for:

1. Finding entities with specific relationships or missing connections

   Example: "Identify all plasmids registered in the last 2 months that do not have linked stable cell lines. Create a table showing: Plasmid ID, Registration Date, Gene Insert, Selection Marker, and Intended Cell Line Type."

2. Condensing complex experimental data into digestible insights

   Example: "Summarize all experiments from Study ST042 conducted in the last 30 days. Include completion status, primary endpoints met (>90% viability threshold), any deviations from protocol, and remaining milestones. Format as an executive summary showing completed vs. planned experiments."

3. Creating structured reports from a template

   [Start by uploading a template file with the desired output structure]

   Example: "Using the attached file ‘Nonclinical Study Report.docx’ as a template, generate a study report for experiment EXP2500001. Fill each section of the template with data from Benchling. Make sure to provide citations for each piece of data."

4. Generating formatted documents from experimental data

   Example: "Generate a non-clinical study report for experiment EXP2500001 including: 1) Study objective and rationale, 2) Test article information with lot numbers, 3) Methods section referencing SOPs, 4) Results with all raw data from linked Results tables, 5) Existing statistical analyses with p-values, 6) Conclusions addressing the primary hypothesis. Include all quality control checkpoints and flag any values outside acceptance criteria (CV > 15%)."

5. Investigating experimental failures and identifying patterns

   Example: "Analyze all failed transfection experiments (efficiency <20%) in Project X. Compare against successful transfections (efficiency >70%) from the same period, examining: cell passage number, DNA concentration, reagent lot numbers, operator, incubation time, and confluence percentage. Create a ranked list of variables most correlated with failure. What caused EXP2500001 to go wrong?"

6. Evaluating data organization and access permissions

   Example: "Audit all notebook entries and entities created in Q4 2024 for compliance issues. Identify: 1) Entries without project associations, 2) Entities missing required metadata fields (lot number, expiration date, creator), 3) Users with edit permissions on archived projects, 4) Sensitive data (patient/clinical) accessible to non-clinical team members. Output a prioritized action list with specific remediation steps."

7. Exploring research directions and identifying knowledge gaps

   Example: "Analyze all data from Project Y to identify research gaps. Map which combinations of our lead compounds and cell lines have been tested, what assays have been run (binding, toxicity, stability), and what concentrations have been evaluated. Create a matrix showing tested vs. untested conditions and recommend the top 5 experiments to run next based on existing trends and missing critical data."

Configuration

While admin configuration is not required, it allows Deep Research to be specialized to the needs of your use case. Without configuration, the Deep Research is aware of your prompt, your prior messages and answers in the same chat, and your Benchling tenant's data model from the perspective of the schemas the user asking the question can access.

Currently, there is one centralized set of guidelines that applies to all Deep Research conversations on your tenant, and you must be a tenant admin to set the configuration. To configure the guidelines:

• Click your profile picture in the bottom-left and select Tenant admin console.
• Click Settings.
• Click AI Settings.
• In the Deep Research guidelines section add any general instructions, abbreviations, definitions, or general information about your organization that might be helpful contextually for any Deep Research or Ask question.
• A Settings changed banner will appear. To apply the changes, click Save.

Here is one example set of guidelines:

Company background:
* Our company focuses on production of novel Antibody therapeutics. 

Common acronyms and abbreviations:
* "STDY" is our internal convention for a research study or related set of experiments. 
 

Changes to the guidelines take effect immediately.

Guidance for writing guidelines

• Guidelines do not need to follow any particular format. Bullet points and sections may help organize the guidelines, but are not required.
• In most cases, the guidelines should be written in the same style as when writing guidelines directly for scientists. Referring to materials you may already have on hand for new hires at your company may be helpful to review as a starting point.
• While the AI system has a broad understanding of scientific topics, it may be helpful to define any relevant terminology or provide additional context. Deep Research has access to the formal names of schemas and entities in your data model, but defining common short names or abbreviations may make it easier for scientists to write more naturally while also increasing the likelihood that Deep Research will make the same assumptions as a colleague would e.g. which of two similarly named schemas is more frequently used. 
   

Risks from AI mistakes

• All AI-based systems may make mistakes.
• The system may produce false positives, which may be misleading.
• Always treat results with appropriate skepticism, and double-check any concerns before relying on the responses or saving them to a Notebook entry or template for broader visibility. 

Limitations

• The AI system is not guaranteed to give the same results when used multiple times, even when run on exactly the same prompt with the same guidelines.
• Guidelines can only be configured for the entire Benchling tenant, and cannot (yet) be configured more specifically for templates, teams, etc.
• The AI system does not have access to information about the use of any other AI-based features in Benchling.

Data Protection & Security

Deep Research operates on behalf of the user who made the query. That means it has access to the same data as the user who made the query, not all of the data within Benchling.

For more information about privacy and security for AI-powered features, see Data protection and security for AI at Benchling.