Agentic Strategy Studio

An opinionated, production-minded starter app that demonstrates advanced Agentic AI patterns in one cohesive workflow:

• Multi-agent orchestration (Planner -> Retriever -> Analyst -> Skeptic -> Synthesizer)
• Hybrid RAG (semantic + lexical retrieval with reciprocal rank fusion)
• Working memory and run history
• Guardrails (PII sanitization + prompt-injection hardening)
• Self-reflection and critique loop
• Evidence-backed answers with source citations
• Lightweight evaluation panel (citation coverage, context precision, response risk)
• Human-in-the-loop controls from the UI

Use Case

The app is focused on strategy and due-diligence intelligence.
You can load internal notes/documents and ask high-stakes business questions.
Agents collaborate to produce an evidence-grounded answer plus concerns and next actions.

Quick Start

pip install -r requirements.txt
set PYTHONPATH=src
streamlit run streamlit_app.py

Optional:

set GEMINI_API_KEY=your_key
set GEMINI_MODEL=gemini-2.0-flash

Without an API key, the app runs in deterministic mock mode for local testing.

Architecture

• src/agentic_studio/rag: ingestion, chunking, and hybrid retrieval
• src/agentic_studio/agents: specialized agents and orchestrator
• src/agentic_studio/core: LLM adapter, guardrails, memory, schemas
• src/agentic_studio/evals: quality and risk metrics
• src/agentic_studio/ui: Streamlit interface

Topics Demonstrated

1. Agent planning and role specialization
2. Tool-using retrieval agent
3. Multi-stage reasoning pipelines
4. Retrieval-grounded generation
5. Defense-in-depth prompt safety
6. Iterative critique and revision
7. Memory and traceability
8. Evaluation and quality gates
9. Human approval touchpoints
10. Extensible model provider abstraction

Next Extensions

• Add SQL and web tools with explicit tool calling
• Add graph-RAG over entities/relationships
• Add automated eval harness with benchmark datasets
• Add async task queue for long-running autonomous jobs