title	emoji	colorFrom	colorTo	sdk	sdk_version	app_file	pinned	license	short_description
AgenticRAG	🚀	indigo	yellow	streamlit	1.50.0	app.py	false	mit	PDF Chat with AgenticAI.

Agentic RAG 🤖

📌 Overview

This project is a agentic AI framework with hybrid RAG system. It demonstrates how to build an LLM-powered agent that can:

• Reason step by step (Plan → Act → Reflect → Answer).
• Use tools (calculator, RAG, internet search, summarization, memory lookup).
• Store and retrieve knowledge with long-term memory.
• Enforce guardrails so the agent stays grounded to documents.

It is designed to be simple, extensible, and easy to plug into applications (e.g., Streamlit chatbots).

Accesible in Hugging Face space - https://huggingface.co/spaces/polojuan/agenticAI

---

🏗️ Architecture

Components

1. LLM Adapter

• Wraps OpenAI (or another provider).
• Handles completions, token limits, and temperature.
• Can be swapped with Anthropic, local models, etc.

2. Agent Core

   • The reasoning loop that implements the ReAct pattern:
     1. Thought → plan next step.
     2. Action → decide which tool to call.
     3. Observation → receive tool output.
     4. Reflection → critique correctness.
     5. Repeat until Final Answer.

3. Tools Registry

   • Each tool has a name, description, JSON schema, and a Python function.
   • Tools include:
   • rag → retrieval-augmented generation
   • internet → limited web search
   • search_memory / write_memory → semantic memory access
   • You can add any other tool required

4. Memory

   • JSONL file for long-term notes.
   • TF-IDF-lite search for retrieval.
   • Stores goals, tool calls, answers, and notes.

5. Critic Module

   • After each step, a “critic” pass reviews correctness and safety.
   • Detects tool misuse, hallucinations, or off-topic answers.
   • Suggests corrections.

---

🔄 Agentic Workflow

Here’s the step-by-step workflow:

1. User Goal: The user submits a query (e.g., “Summarize this article”).

2. Context Building:

   • Agent retrieves relevant memory.
   • Injects available tool manifest.
   • Builds a structured system prompt with rules.

3. LLM Reasoning Loop:

   • LLM outputs:
   • Thought: … → internal reasoning.
   • Action: <tool> {args} → request to use a tool.
   • Final: … → final answer.

4. Tool Execution:

   • Python validates arguments against schema.
   • Executes the tool (e.g., call tool_rag).
   • Returns Observation[...] with results.

5. Reflection:

   • A critic LLM checks if the last step is correct/safe.
   • Flags errors (e.g., wrong tool, hallucination).
   • Suggests adjustments.

6. Loop Control:

   • Continues until:
   • Final: is reached, OR
   • Max steps/tokens hit → fallback answer.

7. Answer Storage:

   • Final answer stored in memory with metadata.
   • Can be retrieved in future queries.

---

🛠️ Setup & Installation

Prerequisites

• Python 3.8+
• Docker (if using containerized mode)
• Access to a language model (OpenAI API key, local LLM, etc.)

Local Setup (without Docker)

git clone https://github.com/palscruz23/agenticAI.git
cd agenticAI

pip install -r requirements.txt

# Run the app (assuming it’s a Streamlit or web frontend)
streamlit run app.py

---

📜 License

This repository is licensed under the MIT License. See the LICENSE file for more details.