Ground answers in project material
The model receives retrieved research context instead of answering Chesapeake Bay questions from general training data.
ArchivedRAG system · 2024
Research internship
SaltCast
At Horn Point Laboratory, I built the chatbot path that turns project material—research papers, salinity data, interviews, and team information—into grounded, streamed answers for SaltCast users.
RAG system
- Retrieval
- FAISS vector search
- Delivery
- Streaming responses over SSE
- Research
- Presented at MIT URTC
01
A public interface to a research project
SaltCast combines hydrologic and oceanographic work around Chesapeake Bay salinity. The chatbot needed to answer questions about both the science and the people behind it without forcing visitors to search across papers, datasets, and project pages.
02
Retrieval before generation
The service embeds the project corpus into FAISS and retrieves relevant material before asking the model to respond. Index options included HNSW and FlatL2 so retrieval behavior could be tuned for the corpus rather than left at a framework default.
03
A conversation, not a blocking request
The FastAPI service streams generated text over Server-Sent Events and maintains bounded conversational context. The browser receives useful output immediately while the longer model operation continues.
Engineering choices
The decisions that shaped it.
The implementation details matter because each one closes a specific failure mode or keeps an important boundary visible.
Tune the index deliberately
FAISS index choices are explicit, allowing latency and retrieval quality to be evaluated against the actual corpus.
Stream the slow boundary
Server-Sent Events keep the interface responsive while generation is still in progress.
Where it landed
A research assistant that shipped inside SaltCast and was presented as part of the work at the MIT Undergraduate Research Technology Conference. The original product surface is no longer maintained, so this page preserves the work as an archived case study.
Next case study
Testif.AI