CCC is pleased to announce the beta launch of a generative AI (“genAI” for short) enabled Chat Mode for RightFind Navigate, a milestone in achieving our vision of enabling researchers to have a natural language interface for scientific discovery that speeds their time to knowledge but, crucially, recognizes the responsibility to build trust in AI systems.
Chat Mode in RightFind Navigate provides:
- Natural language interface for conducting search with a response generated by an LLM based on the top five results for the query
- Conversational context that allows the user to ask follow-up questions based on previous answers
- Integration with Navigate core features including “add to library” and “get content” from RightFind to obtain the full text
- A full list of search results based on the query so users can see the sources for the answers the system produces
Why Chat Mode?
The choice to introduce a genAI Chat Mode in RightFind Navigate was a careful one. CCC already has a strong track record in applying advanced techniques such as machine learning models, knowledge graphs, and named entity recognition to enhance scientific discovery and facilitate exploration of concepts using knowledge graphs, but we did not want to rush to market and pile on as yet another vendor seeking AI buzz without any consideration of user needs. There are significant challenges in genAI implementations, and important considerations related to copyright, and we know that our clients expect us to approach such an implementation in a thoughtful manner.
Through iterative discussions with our clients, we identified use cases that we could help with. A central one is the agitation that researchers face when they see a search box in (yet another) discovery tool. RightFind Navigate is extremely powerful in its integration of disparate public and licensed data sources and creation of a personalized linked data layer, but we know that end users still face the hurdle of crafting the right search syntax to develop the most effective query to match their question.
It was this challenge that we tailored our beta release to address: users can interact in natural language and receive a narrative response that enables them to access a broader set of results relevant to their topic. A great example of this is a researcher in early target identification who can have a conversation with the system:
- I want to know everything about Disease X
- What is it linked to?
- Are there relevant cell models?
- What companies are already working on this?
For each response to the above, the user can earmark a set of articles to dig deeper, learn more, and validate their understanding.
Responsibility and AI systems
CCC has been an industry leader in the discussions about genAI and copyrighted materials, having introduced licensing for internal AI usage. Our approach to introducing AI-enabled features keeps these responsibilities top of mind. This beta release of Chat Mode takes a retrieval augmented generation (RAG) approach that uses PubMed abstracts/citations and Open Access CC BY full-text literature for grounding the response of the system.
In upcoming releases, a rights awareness layer will incorporate additional full-text scholarly literature that clients have access to and is covered by the new internal AI usage rights under CCC’s Annual Copyright License, enriching the experience for CCC licensees.
Limitations
Any system has limitations, and AI systems are no different. We understand that many of our users are fatigued by the generic disclaimers seen in general purpose genAI chat systems to “check the facts.” Human validation remains a vital aspect of good science, and the linking of the chat narrative response to underlying literature supports this practice.
To reduce the risks in certain areas, CCC filters the LLM responses in particular areas to avoid providing questionable information. For example:
- Prompts looking for people-related information (e.g., “Make a list of the top five researchers in x”)
- Prompts asking for qualitative opinions (e.g., “What is the best weight-loss drug?”)
- Prompts asking for assistant-type actions (e.g., “Make a list of the top articles in x domain, then create a library and populate it with these articles; share it with colleagues a, b, and c”)
Participants who ask these types of questions will see an explanation that the prompt is outside the scope of what the chat is designed for. We envision exploring ways to provide support in these areas, as we conduct more investigation, research, and development on these types of requests.
Future plans
With a guiding vision to provide a robust genAI experience in a workflow where scholarly research is primary, and where clients expect a built-in respect for copyright, we are excited to see how chat mode will evolve. We will create the roadmap in close consultation with our clients and have already conducted several other beta deployments; these have enabled us to collect real-world feedback (in a privacy preserving manner), monitor prompt techniques and responses, clarify user understanding, and compile a backlog of improvements to make the system even better for our users.