OpenScout

OpenScout is an open-source Agentic Retrieval-Augmented Generation (RAG) system built using the LangGraph framework. It intelligently routes user queries between a local vectorstore and real-time web search, ensuring optimal relevance, freshness, and efficiency.

OpenScout leverages open-source LLMs for both routing and generation, enabling a transparent, controllable, and vendor-neutral AI stack suitable for production and research use cases.

---

Core Capabilities

• Agentic Query Routing with LangGraph Deterministic, graph-based agent orchestration using LangGraph for explicit and debuggable control flow.

• Dual-Model Architecture

  • Router Model: qwen3:8b for lightweight, fast decision-making
  • Generator Model: gpt-oss:20b for high-quality response generation

• Local-First RAG Strategy Queries are routed to the local vectorstore whenever possible, preserving privacy and reducing latency.

• Real-Time Web Search via Tavily Automatically invoked for queries requiring fresh or external information.

• Open Embedding Stack Uses embeddinggemma:300m for efficient and high-quality semantic embeddings.

• Modular and Extensible Each component i.e., routing, retrieval, search, and generation is independently swappable.

---

System Architecture

OpenScout is implemented as a LangGraph-powered agentic workflow, where each node represents a discrete reasoning or retrieval step.

High-Level Flow

1. User submits a query

2. Router agent (qwen3:8b) evaluates intent and freshness requirements

3. Query is routed to:

   • Local vectorstore retrieval, or
   • Tavily-powered live web search

4. Retrieved context is normalized and aggregated

5. Generator model (gpt-oss:20b) produces the final answer

---

LangGraph ensures:

• Explicit state transitions
• Deterministic routing logic
• Full traceability and debugging support

---

Technology Stack

Agent Framework

• LangGraph: Graph-based agent orchestration and control flow

Language Models

• Router: qwen3:8b
• Generator: gpt-oss:20b

Embeddings

• embeddinggemma:300m

Vectorstore

• SKLearnVectorStore

Web Search

• Tavily API (real-time, LLM-optimized search)

Evaluation

• LLM as a Judge implemetation to check for Hallusinations and Answer correctness

---

Example Queries

Internal Knowledge Query

"What is CoT prompting?"

→ Routed to vectorstore

Fresh / External Query

"What are the latest developments in the EU AI Act?"

→ Routed to Tavily web search

---

LangGraph Agent Design

Router Node

• Powered by qwen3:8b

• Classifies:

  • Knowledge domain (internal vs external)
  • Freshness requirements
  • Confidence threshold for local retrieval

Retrieval Nodes

• Vectorstore retriever
• Tavily search retriever

Generation Node

• gpt-oss:20b
• Receives structured context and provenance metadata

---

Use Cases

• Enterprise internal knowledge assistants
• Research copilots with real-time awareness
• Privacy-sensitive RAG systems
• Local-first AI assistants with web fallback
• Developer documentation bots

---