Getting Started: Domain Scientists

Getting Started: Domain Scientists

This guide helps domain scientists (biologists, chemists, data scientists) get started with ADEPT for scientific workflows.

---

Prerequisites

• Docker Engine 24+ with Compose v2
• 16GB RAM minimum (32GB recommended)
• LLM API keys (Azure OpenAI, AWS Bedrock, or Anthropic)

---

Quick Start

# Clone and configure
git clone https://github.com/pnnl/adept-agentic.git
cd adept-agentic
cp .env.example .env
# Edit .env with your LLM API keys

# Start the platform
make start

# Verify services are healthy
make validate-service-health

---

Accessing ADEPT

Streamlit Chat Interface

Navigate to https://localhost/streamlit/ after login (Keycloak SSO).

Use natural language to:

• Run BLAST sequence alignments
• Search PubChem for molecular properties
• Execute Python/R code in the sandbox
• Upload and process CSV/PDF files with RAG

JupyterLab

Navigate to https://localhost/jupyter/ for interactive notebook development.

17 tutorial notebooks cover setup, end-user workflows, developer reference, and operations.

CLI (afk)

cd examples/adept_connectors
uv sync
uv run afk chat "What tools are available?"

---

Available Scientific Tools

Category	Tools
Bioinformatics	BLAST sequence alignment, UniProt protein database, AlphaFold structure prediction
Chemistry	PubChem compound search, molecular property calculation
Data Processing	CSV/PDF RAG with ChromaDB, batch file operations, SQL transformation
Code Execution	Sandboxed Python/R execution with import validation
Web Research	Multi-engine web search, web scraping
File Management	Upload, download, presigned URLs, session-scoped directories

---

Example Workflows

Protein Analysis

You: "Run a BLAST search for the sequence MKTAYIAKQRQISFVKSHFSRQLEERLGLIEVQAPILSRVGDGTQDNLSGAEKLV against the nr database"

ADEPT: [Runs BLAST tool, returns alignment results with E-values and identity scores]

You: "Now look up the top hit in UniProt and get its function annotation"

ADEPT: [Queries UniProt, returns protein function, GO terms, and pathway information]

Data Analysis

You: "Upload my experiment results and create a RAG index"

ADEPT: [Creates session-scoped vector index from uploaded CSV/PDF files]

You: "What are the top 5 compounds by binding affinity?"

ADEPT: [Queries RAG index, returns structured answer with source citations]

Code Execution

You: "Write and run Python code to plot a histogram of the binding affinities from my uploaded data"

ADEPT: [Generates code, executes in sandbox, returns plot as base64 image]

---

RAG Workflows

ADEPT supports visibility-scoped RAG indexes:

Scope	Prefix	Access
Session	s_	Current session only (default)
User	u_	Persists across sessions for you
World	w_	Accessible to all authenticated users

Processing types are automatically detected:

• Auto: Content analysis determines processing strategy
• RAG: Vector embeddings for semantic search
• SQL: Structured data for SQL queries
• DataFrame: Tabular operations

---

Multi-Agent Teams

Request specialized teams for complex tasks:

You: "Create a team with a biologist and data scientist to analyze my proteomics dataset"

ADEPT: [Creates multi-agent session with domain-specific LLM routing]
  - biologist: Uses biology-optimized model
  - data_scientist: Uses data analysis model

Each worker agent has access to all scientific tools with session isolation.

---

Further Reading

• Architecture Overview
• RAG Workflows
• ADEPT Connectors
• Onboarding guide: docs/DOMAIN_SME_VIBE_CODING_GUIDE.md (in repository)