Frequently Asked Questions

What's the benefit of using tundra's disease-scoped databases?

Our databases constrain and deepen the space where your AI models search. For highly-specific research tasks, databases from tundra provide your model with relevant, 250+ recent full-text scientific papers and patents, whereas Claude or GPT in stock form would refer to titles and abstracts it can find through a shallow, 5–10 paper web search.

An ALS researcher had this to say regarding the benefits to researchers new to their field:

"It's really good. For ALS, the answers to my questions were pretty much exactly where the field is at now… Overall I think the answers it gave me are great, in part because the library was probably curated for exactly what I wanted to do. It doesn't give me any new info per se, but it reinforces my interpretation of the field. I could see how this could be really useful for someone who isn't an expert in the field."

What are some specific use cases that your early users have revealed to you?

Generation of in vitro and in vivo preclinical experiment lists: A user was interested in looking through disease-specific literature to identify all the possible experiments that they could run over the course of H2L and LO stages of their drug discovery program. With tundra, they were able to go from a specific prompt to a list of experiments, each containing a PMC ID and quote, in roughly 5 minutes. Sample conversation

Experimental design landscaping: A user was interested in learning about ways that they could optimize the design of a pathology experiment they were developing for a specific disease area. Using tundra, they found the range of specimen thickness, specimen types, and antibody labeling strategies that would have been impossible to uncover using traditional AI search. Sample conversation

Using a 'digital researcher' to validate a hypothesis: A user was interested in receiving guidance regarding their early discovery strategy, and used our 'Bob Langer' database to learn about recent trends on mRNA-related design and delivery strategies. We imagine users creating 'digital research teams' in the future to validate / hone scientific ideas. Sample conversation

How is this different from adding downloaded papers into Claude?

When you upload a stack of papers (typically in PDF form, added directly to the chat box), your model's memory (or context window) is immediately partially filled with all of the information you provided. When you ask a new question, your model will sort through the entire memory, which degrades the quality of responses. More importantly, when papers are stored this way there is no structured mechanism in place to ensure that models can faithfully cite and quote the papers that they depended on to generate their answers. Tundra is the only way to guarantee that claims produced by AI are backed up by real, relevant scientific sources.

How is this different from adding URLs to websites into Claude?

When you add a URL to Claude, the model attempts to fetch web content using browser calls. In the case that it can retrieve a full source, it typically looks at shorter, truncated portions of papers. When web search fails (due to paywalls or webpage security), your answers will be derived from the model's weights, which typically recall shallow information like titles and abstracts. As a result, answers lack depth, suffer from poor provenance, or may hallucinate. Our advice to users is that uploading a stack of downloaded PDFs to Claude is better than inputting URLs. Of course, the best solution is to plug in a domain-specific tundra database instead!

How is this different from the existing PubMed connector made by Anthropic?

The main difference is that PubMed indexes abstracts from paywalled articles while we do not. This means that answers generated with the PubMed connector activated may rest on shaky or incomplete foundations. On the other hand, tundra databases only include full-text open access publications, and models are prohibited from quoting titles or abstracts. As a result, we guarantee that answers generated from our databases can be traced back to core technical details that the PubMed connector inconsistently retrieves. To prove this, we've included a sample conversation comparing answers generated with our tundra-EGFR connector versus the PubMed connector.

You claim you don't store any of my data. How can I be sure of that?

Your research questions never reach our servers. We use Cloudflare Workers to host our MCP servers, and as a result, only tool calls and keywords registered by your AI model are sent to the MCP, to perform lookup tasks in our Cloudflare R2 databases. The AI model provider you choose to use will save your prompts for model training, and the output generated using our MCP servers, but the information directly sent to our MCP servers throughout this process is fully wiped at the end of each interaction. For more information about the security policies of Cloudflare for R2, Workers, and their general policies, please visit the respectively attached links.

How can I set up tundra's disease-scoped databases in Claude or GPT?

1. Go to Settings in the bottom left of the Claude chat.
2. Click on Connectors.
3. Click on Customize.
4. Click the 'plus' button.
5. Select Add Custom Connectors.
6. Enter your MCP URL (tundra-bookshelves etc.) and name it as you want.
7. Make sure to allow all permissions! Refresh your Claude chat window and you're ready!