Seeker

A distributed, event-driven agentic framework designed to execute complex workflows using a hybrid planning system: a planning-heavy cold-path (LLM-generated plans) and a cached fast-path (pre-learned skill-graphs). It continuously refines its behavior by mining executed traces into reusable skill definitions.

Tech Stack

• Go (Golang): Core programming language
• NATS JetStream: High-performance messaging, event streaming, and persistent KV state store
• MongoDB: Document database for persisting claims, entities, and sources (Knowledge Graph)
• Qdrant: Vector database for semantic skill matching and recall
• SearxNG: Search engine for web search (open-source alternative to Google Search)
• Chromium: Headless browser for web scraping and browser-based tasks

Critic Agent Confidence Assessment

The Critic Agent assesses the reliability of synthesized claims using a weighted harmonic mean of independent source verifications. The confidence score $C$ is calculated as:

$$C = \frac{2 \times (W_s \times W_v)}{W_s + W_v}$$

Where:

• $W_s$ is the semantic similarity of the source.
• $W_v$ is the cross-validation agreement metric.

This thresholding determines whether the Human-in-the-Loop (HITL) fallback is invoked.