Experimental Sandbox

Agentic
RAG Sandbox

Build, test, and reason with your data using agentic retrieval and tools. Extend your assistant with MCP—not prompts.

Local
Embeddings
MCP
Tool Protocol
RLS
Security

What are the key findings from the research paper?

Based on the retrieved documents, the key findings include three main areas of improvement...

confidence: 0.94research.pdf:12
Used: vector_search, planner

Most RAG demos break in real usage.

Generic retrieval systems fail when they hit production complexity. Agentic RAG Sandbox was built to fix these problems at the system level.

Static retrieval with poor recall
Hallucinations when context is weak
No ordering or confidence signals
No tools beyond document search
No isolation for personal data
Capabilities

What you can do with the sandbox

01

Intelligent Data Ingestion

Upload PDF, DOCX, CSV, JSON, and text files. Context-aware chunking preserves meaning with local embeddings via transformers.js.

02

High Quality Retrieval

Adaptive vector search with fallback logic. Ordered context reconstruction with confidence-aware injection and source attribution.

03

Agentic Tool Use

Planner-driven decision making with MCP-based tool discovery. Clean separation between reasoning and execution.

04

Secure & Personal

Email/password auth with strict row-level security. Each user's data stays completely isolated.

How the system thinks

Instead of forcing every query through retrieval, the system plans first, retrieves only when needed, and uses tools when documents aren't enough.

1

Query

User sends a question

2

Plan

System decides approach

3

Retrieve

Fetch relevant context

4

Execute

Run MCP tools if needed

5

Respond

Grounded, cited answer

Planning First

The planner evaluates each query to decide if retrieval, tools, or direct response is best.

Confidence Scoring

Context is assembled with confidence scores, allowing the LLM to weight information appropriately.

Ordered Context

Retrieved chunks are ordered for coherent narrative flow, not random context stuffing.

More than a chatbot

Traditional RAG

Agentic RAG Sandbox

Always retrieves
Plans before acting
Blind context stuffing
Ordered and scored context
No tools
MCP based tools
Hallucination prone
Refusal when info is missing
Stateless
Persistent personal data

Built for people who care about systems

Not a consumer toy. A learning and prototyping environment.

Developers

Building RAG or agent pipelines

Researchers

Experimenting with retrieval

Students

Learning LLM architecture

Builders

Prototyping assistants

Tech Stack

Built with modern, inspectable tech

SupabaseAuth, storage, pgvector
transformers.jsLocal embeddings
GroqFast inference
MCPTool interfaces
Next.jsFrontend

Client → Edge Functions → Vector DB → LLM → Response

Experiment with agentic RAG the right way.

A transparent, secure, and extensible environment for building retrieval augmented generation systems.

Launch Sandbox