Ask Susie: building an AI assistant that knows what it doesn't know

The midnight problem

Wildlife rehabilitation does not run on business hours. A volunteer gets a call at 11pm about an injured box turtle found on a road. They bring the animal in and need to make quick decisions: is this an emergency that needs an emergency vet or a more senior rehabber tonight, or can it wait until morning? What temperature should the enclosure be? Is the animal in enough distress to warrant immediate intervention?

The answers to those questions exist. We have built up care protocols, species guides, and rehabilitation records over years of work, all living in a shared Google Drive. But at 11pm, searching through folders while also managing an injured animal is not a realistic ask. Knowledge that is hard to access in the moment is not as useful as it should be.

That is the problem Ask Susie was built to solve.

Meet Susie

Ask Susie is an AI-powered chatbot named for one of our Eastern Box Turtle ambassadors. Volunteers and rehabbers can type a question in plain language. Something like "what temperature should a box turtle enclosure be during overwintering?" or "what are the signs of respiratory infection in a snapping turtle?" gets a direct, sourced answer in seconds.

What Susie is not is a general-purpose AI. She is not ChatGPT. She does not pull answers from the internet at large, invent plausible-sounding protocols, or synthesize information from sources we have not reviewed. What she knows, she knows because we taught her, and she is designed to tell you clearly when she does not know something.

Why grounding matters, and what it actually means

You may have heard the term "hallucination" used to describe AI. It refers to a model confidently generating information that sounds correct but is simply fabricated. General AI assistants do this because they are trained to produce fluent, helpful-sounding responses, and sometimes that fluency outpaces accuracy.

In most contexts, hallucination is an inconvenience. In wildlife care, it is dangerous. An AI that invents a medication dosage, misremembers a temperature range, or confuses the care needs of two similar species could directly harm an animal. This is non-negotiable for us: the AI must only tell our volunteers things we can stand behind.

The solution is a technique called retrieval-augmented generation, or RAG. Instead of relying solely on what the model was trained on, RAG adds a step: before answering, the system retrieves the most relevant documents from a curated knowledge base and uses those as the actual source of the answer. The model's job is to read those documents and synthesize a clear response, rather than generating information from memory.

For Susie, that shelf is our Google Drive. Every care protocol, species guide, and intake procedure we have written lives there. When a volunteer asks a question, Susie searches that library first. If the answer is there, she finds it and surfaces it. If it is not, she says so, and that admission is just as valuable as the answer would have been.

How it's built

For those who want the technical picture: Ask Susie is built on Google's Agent Development Kit (ADK) with Gemini as the underlying language model, hosted on Google Cloud Run. The knowledge base is indexed in Vertex AI Search, which handles the retrieval side of RAG. It finds the most relevant chunks of our documentation given a volunteer's question and passes them to Gemini to synthesize the response.

A Cloud Function runs on a daily schedule to sync our Google Drive documents into Cloud Storage, so the knowledge base stays current as we update and add protocols. When we improve a care guide or add a new species document, Susie knows about it the next day without any manual steps.

For questions that fall outside our knowledge base (natural history, species identification, general biology), Susie has access to Google Search as a second tool. The system routes between the two depending on the nature of the question: care and rehabilitation protocols come from our documents, everything else can draw on the broader web. Cases that require actual veterinary judgment are directed to our advisory veterinarian.

1. 1
   Volunteer asks a question. Plain language, no special formatting or commands required.
2. 2
   Tool selection. The agent decides whether to search SERC's knowledge base, run a Google Search, or both, based on the type of question.
3. 3
   Retrieval. Vertex AI Search finds the most relevant sections of our care protocols and species guides.
4. 4
   Response generation. Gemini synthesizes a clear answer grounded in the retrieved documents.
5. 5
   Logging. Every exchange is logged as a potential knowledge gap for volunteer review.

The knowledge gap loop

One of the features we are most proud of is not the AI at all. It is the feedback loop built around it. Every question a volunteer asks is logged for review, and volunteers can rate responses with a simple thumbs up or down. The admin dashboard surfaces questions that went unanswered or got poor ratings, which tells us exactly where our documentation has gaps.

In other words, every time Susie says "I don't know," that is not a failure. It is a signal. It tells us what to write next. The more the chatbot is used, the better our underlying knowledge base gets, and the more useful the chatbot becomes. It is a loop that makes the whole system smarter over time without any AI involved in that improvement at all: just people identifying gaps and filling them.

What this means for other rehabbers

Building Ask Susie required technical infrastructure that most small rehabilitation programs do not have in-house: cloud hosting, AI APIs, custom development. We are not suggesting every program needs to build their own version. But the principle behind it is worth thinking about regardless of your tech situation.

Ask Susie is only as good as the documentation we gave it. That documentation (the care protocols, the species guides, the intake procedures) would be valuable on its own even without any AI layer at all. Getting that knowledge out of people's heads and into a shared, searchable format is step one, with or without a chatbot on top of it.

If you are curious about how Ask Susie works or want to explore building something similar for your own program, the source code is open and available on GitHub as ask-serc. We are happy to talk through what we built and what we learned.