The full source code is available on GitHub: vero-code/biotessera

What it does:
Biotessera acts as an intelligent research assistant that allows users to ask complex questions about NASA's space biology research and receive synthesized, factual answers based on a knowledge base of hundreds of scientific publications. It can also perform live searches on external NASA databases like the Open Science Data Repository (OSDR) to enrich its answers with links to raw datasets.

How it works:
The project is built on a multi-agent architecture (Coordinator/Worker pattern):

Data Preparation: An offline pipeline processes hundreds of full-text publications, extracting metadata and generating vector embeddings for every text fragment (a "tessera"). These are stored in a local ChromaDB vector database (TesseraStore).
Retrieval: The main agent, TesseraConductor, receives a user's question and delegates tasks to specialized tools (TesseraMiner, DataFinder).
Synthesis: The TesseraConductor gathers all retrieved information and uses the Google Gemini LLM to generate a single, coherent, and sourced answer.

Benefits & Intended Impact:
The intended impact of Biotessera is to accelerate scientific discovery and support future mission planning. Its key benefits include:

Accelerated Research: Provides researchers with synthesized answers from hundreds of papers in seconds, not weeks.
Increased Trust & Verifiability: Every statement generated by the AI is directly traceable to a source publication, preventing "hallucinations" and building trust in the results.
Discovery of Novel Connections: By analyzing a vast corpus of data, the agent can help surface non-obvious connections between different studies and experiments.
Actionable Data: By linking publication findings to raw datasets in the OSDR, it bridges the gap between research and the underlying data, making it more actionable.

How is it creative?
The creativity of Biotessera lies in its architectural approach:

Agentic System: This is not just a search engine. It's an AI "thinking" system that can understand complex queries, break them down into steps, and decide which tool to use, mimicking the workflow of a human researcher.
The "Tessera" Metaphor: In Biotessera, knowledge is treated as monolithic documents, but as modular "tesserae" (tiles of information). This allows the agent to build answers like a mosaic, combining verified pieces of information into a new, comprehensive picture.
Multi-Source Synthesis: The ability to dynamically combine information from a static, internal knowledge base (TesseraStore) with real-time data from external sources (DataFinder) is a powerful and innovative feature.

Factors considered:
As a solo participant, key factors in the design were feasibility and focus. The project was scoped to be achievable within 48 hours, prioritizing a robust core functionality (the agent and TesseraMiner) over adding too many features. The choice of Streamlit for the UI was a strategic decision to enable rapid prototyping, allowing more time to be spent on the complex AI backend. The entire architecture is built to be modular and scalable, with a strong emphasis on providing trustworthy, source-backed answers.

What tools were used:

Backend: Python
AI/ML Framework: LangChain
LLM & Embeddings: Google Gemini
Vector Database: ChromaDB
Hugging Face Datasets: vero-code/biotessera-database
UI: Streamlit
Data Processing: Pandas, BeautifulSoup

2025 NASA Space Apps Challenge

Team Biotessera (Veronika)