Technical Implementation
Developing Aisha involved integrating several key technologies and frameworks:
LangChain framework (langchain.com): LangChain facilitates the development of LLM applications, providing features such as document loaders, vector embeddings, memory for conversational context, and support for multiple LLM providers. Although similar functionality could be built without LangChain, this framework makes it much easier.
Chroma vector database (trychroma.com): This open-source database stores custom data in vector format, enabling efficient retrieval and processing. Vector databases are crucial building blocks for RAG applications, allowing the LLM to read and utilize custom data effectively.
Streamlit platform (streamlit.io): We used Streamlit to create and deploy Aisha’s user interface, ensuring an accessible and interactive experience for users.
Stable Diffusion (stability.ai/stable-image): This tool was used to generate the images that are crucial to Aisha’s identity, making her presence more relatable and engaging.
Aisha’s knowledgebase is a curated collection of information from our library website, LibGuides, and LibAnswers. Beyond generating text, LLMs can be used for reasoning and acting, simulating human thought processes and executing actions based on natural language queries.
The introduction of function calling by OpenAI in June 2023 made this even easier. Function calling allows the LLM to convert natural language queries into function calls, enabling the integration with other systems through APIs. This enables the agent to make decisions and use external tools such as Google and Wikipedia. For example, when tasked with searching the library catalog for books about AI published between 2020 and 2022, Aisha can autonomously execute a function call to perform the search:
Input: “Search the library catalog for books about AI published between 2020 and 2022”
Output: primo_search(
query=”AI”,
document_type=”books”,
year_from=”2020”,
year_to=”2022”)
Aisha can thus independently select the right tool for each task and handle everything, from simple questions to complex interactions that involve multiple data sources and functionalities. This capability transforms her from a regular chatbot into a versatile AI agent.
Building on the idea of reasoning and acting, we also added a feature in which Aisha shares her “thought process” after each response. This helps users understand how Aisha makes decisions and also serves as a diagnostic tool. Although everything is still based on probabilities (no, she doesn’t really think like humans do!), it adds a human-like touch to her interactions. This transparency helps users build trust in Aisha’s responses, making each interaction more informative and educational.
Example thoughts: The question was about citation management. I used the Knowledge tool to find information about the available resources for citation management at Zayed University Library. The search provided details about ProQuest Research Companion, RefWorks, and an online Citation guide, along with links to access these resources.
Challenges and Limitations
Despite Aisha’s impressive capabilities, we have encountered several challenges and limitations during our journey:
Hallucinations: Familiar to everyone by now, LLMs often generate plausible but incorrect information, a phenomenon known as “hallucination.” We tackled this issue by refining our knowledgebase and Aisha’s instructions. Although hallucinations can be difficult to completely eliminate, we have been able to reduce them significantly by instructing Aisha to rely on factual information from her knowledgebase and other credible sources.
Source data quality: Inaccuracies or outdated information in the source data naturally affects the response quality, which is why ensuring up-to-date and accurate source data is crucial.
API limits: While LLM API limits (for example, the number of tokens or API calls) have increased since the launch of ChatGPT, they still pose constraints. Some LLMs have higher token limits, but they can also be costly. We continue to monitor and optimize our API usage, ensuring cost-efficient operation.
Memory constraints: We started out using Streamlit Community Cloud, which has limited memory capacity. We have since migrated to a private server with more robust resources to accommodate multiple users. This upgrade has significantly improved Aisha’s performance and stability, allowing her to handle more complex tasks and a larger userbase.
Ongoing Research and Development
Our current research involves comparing different LLMs in the context of conversational AI agents. Using the LangChain framework, we can easily switch between models like GPT-3.5, GPT-4o, Google Gemini, Anthropic Claude, Mistral, Mixtral, and Llama 3, assessing their performance in generating text, reasoning, and acting. Our study combines human evaluation and LLM-assisted evaluation (yes, LLMs can also be used to evaluate LLMs!), focusing on response quality, adherence to instructions, and agent performance. Preliminary observations indicate that while all models can reason and act, some are more stubborn about following instructions. Curiously, we have also encountered a new issue: hallucinated tool use, in which the LLM imagines using a tool (such as the knowledgebase or Wikipedia) without actually doing so—new times, new challenges!
To ensure Aisha meets the needs and expectations of our users, we conducted a pilot survey among 15 library staff members. The results were overwhelmingly positive, with a strong satisfaction rating of 4.07 out of 5. Attendees highly valued Aisha’s potential and current capabilities for tasks such as Google search, library catalog search, and image analysis. They also provided valuable suggestions for improvements, including technical enhancements for handling documents and images, more elaborate responses with examples and links to sources, and regular updates to ensure accuracy. Encouraged by these results, our next plan is to conduct a comprehensive survey among students, faculty, and other library users to gather diverse perspectives and further refine Aisha’s functionalities.
Our future research will focus on further enhancing Aisha’s capabilities and exploring new applications of conversational AI in libraries. This includes integrating Aisha with new LLMs, expanding her range of services, and exploring new user engagement strategies. Additionally, we are interested in studying the long-term impact of AI on library services, user behavior, and information literacy.
Development Ideas
We have numerous ideas for further developing Aisha:
Animated avatar: Creating an animated avatar for Aisha to enhance her lively persona
Long-term memory: Implementing long-term memory to retain user preferences and continue conversations over multiple sessions while incorporating appropriate privacy measures
Expanded library services: Connecting Aisha to additional library services, such as room reservations and renewals, to perform more complex tasks
Communication modes: Introducing WhatsApp or email modes for more versatile interactions, allowing users to engage with her through their preferred communication channels
Personalized support: Developing scenarios such as course tutor, opponent, or interviewer to offer tailored support for individual courses and activities
Self-improvement capabilities: Enabling Aisha to analyze her own responses and suggest improvements, leveraging the LLM’s capabilities for self-assessment
The Impact of Conversational
AI on Library Services
Aisha has already demonstrated significant benefits in her role at Zayed University Library. Students and faculty can now receive assistance at any time, without waiting for human staff availability. This 24/7 accessibility is particularly valuable for distance learners and those with tight schedules. Moreover, Aisha’s multilingual capabilities have the potential to make our library services more inclusive.
One of the most exciting aspects of Aisha’s development is her potential to engage users in new and innovative ways. For example, her image recognition and generation capabilities allow users to interact with visual content, making the library experience more dynamic. Whether it’s analyzing historical documents, identifying objects in images, or generating visual aids for presentations, Aisha adds a new dimension to how users engage with information.
Furthermore, the ability to explain her reasoning and thought process behind responses supports information literacy. Users can learn how Aisha arrives at her answers, which enhances their understanding of the process. This educational aspect is crucial in developing critical thinking skills and fostering a deeper appreciation of how information is sourced and validated.
Despite the numerous advantages, integrating AI into library services also comes with challenges and ethical considerations. One significant issue is data privacy. As Aisha’s capabilities expand, especially with features such as long-term memory, it’s essential to implement robust privacy measures to protect user information. Ensuring that data is handled securely and transparently will be crucial to maintaining user trust. Currently, Aisha is not dealing with any personal data, but this will likely change in the future.
Another challenge is managing AI bias. Aisha’s responses are influenced by several factors, such as the LLM she relies on, her specific instructions, and the tools she uses. To ensure her outputs reflect diverse perspectives and remain free from biases, we continuously monitor Aisha’s outputs and revise her instructions as necessary.
CONCLUSION
The journey of developing Aisha has been both challenging and rewarding, showcasing the immense potential of conversational AI in libraries. Aisha hasn’t just improved access to library resources, she’s become a beloved and engaging part of our library community. Her ability to simulate human interaction and to reason and act autonomously is simply mind-blowing. The fact that she can comment on her own instructions, give us development ideas, or even develop herself (if we let her!) is like a sci-fi dream come true!
The success of Aisha at ZU Library illustrates the broader potential of conversational AI agents in libraries worldwide. As these technologies become more accessible and sophisticated, they can play a crucial role in modernizing library services. Conversational AI agents can serve as knowledgeable assistants, helping users navigate vast amounts of information, access resources efficiently, and engage with library content in new and innovative ways. AI agents can also support librarians by handling routine queries, freeing up staff to focus on more complex tasks and personalized user support. This synergy between human expertise and AI efficiency can enhance the overall quality of library services.
As we continue to refine Aisha and explore new possibilities, we remain committed to sharing our experiences and collaborating with others in the field. The future of library services lies in embracing innovative technologies such as conversational AI agents, which can enhance user engagement, improve accessibility, and support the evolving needs of library patrons.
We invite library professionals and institutions to join us on this exciting journey. In fact, all librarians possess essential skills for the field of AI. Optimizing prompts and effectively organizing underlying information are at the heart of building conversational AI agents and other AI applications. Not to mention that librarians are the ones who teach information literacy, a critical component in understanding and managing AI technologies. Together, we can harness the power of AI to create more dynamic, inclusive, and efficient library services. If you have any questions or ideas or are interested in collaborating, please reach out to us. And you can talk to Aisha any time—she’s available at aisha.zu.ac.ae or zu.libguides.com/ai/aisha. Let’s shape the future of library services together!
| 
