benchen4395/alpha_agentic_search

Alpha Agentic Search — 分层记忆 RAG 检索问答系统

What it solves

Alpha Agentic Search (AAS) addresses the instability, high latency, and lack of personalization in traditional AI search agents. It provides a controllable, low-latency framework for web-based Q&A that improves over time by implementing a hierarchical memory system, ensuring that frequent or previously answered questions are handled with millisecond response times while maintaining high factual accuracy.

How it works

The system follows a classic Route → Rewrite → Retrieve → Verify → Summary pipeline:

  1. Short-circuiting (L1): Checks a QA cache for exact or fuzzy matches to return answers instantly.
  2. Routing: Determines if a specialized tool (e.g., Weather, GitHub, arXiv) can answer the query; otherwise, it proceeds to retrieval.
  3. Rewriting: Injects time and location context into the query to improve search precision.
  4. Hierarchical RAG (L1–L5): Parallelly retrieves data from five layers: L1 (QA Cache), L2 (Wikipedia), L3 (History), L4 (Real-time Web), and L5 (Knowledge Graph). Results are fused using Reciprocal Rank Fusion (RRF) and calibrated to a probability of relevance.
  5. Summarization: An LLM synthesizes the retrieved evidence into a final answer with source attribution and citation verification.

Who it’s for

  • AI Engineers and Architects looking for a production-ready search agent framework that prioritizes stability and latency over pure research.
  • Developers wanting to build personal search assistants with a "get stronger with use" memory mechanism.
  • Beginners in the AI search space who want a reference implementation of a multi-agent search pipeline.

Highlights

  • 5-Layer Memory Stack: Combines fast caches, textbook commonsense, user history, real-time web, and structured knowledge graphs.
  • Cross-Layer Score Calibration: Uses Platt scaling to map different scoring metrics (cosine, rank, etc.) into a unified confidence score.
  • L1 Cache Guardrails: Implements a strict admission policy and "slot gate" consistency checks to prevent the caching of wrong or outdated answers.
  • Prompt Injection Protection: Employs a three-layer defense (content cleaning, XML structural delimiters, and system prompt guards) to secure untrusted web content.
  • Latency Governance: Features layered latency budgets, soft timeouts, and a strategic "RAG-first, LLM-last" warmup sequence to minimize TTFT.
  • Entity Disambiguation: Uses a combination of popularity (in-degree) and query-context signals to resolve ambiguous entities in the Knowledge Graph.

Related

  • Project
    sopaco/cortex-mem

    A Rust-based long-term memory framework for AI agents that uses a three-tier hierarchical architecture and a virtual filesystem to provide persistent, context-aware interactions.

  • Project
    aigc-apps/PAI-RAG

    An open-source Agentic RAG framework that combines multimodal understanding, ReAct-based agents, and enterprise-grade knowledge management for complex intelligent Q&A.

  • Project
    java-up-up/nexus-agent

    An enterprise-grade AI agent platform that implements a complete closed-loop of document governance, hybrid RAG retrieval, and ReAct-based autonomous agents for production environments.

  • Project
    GiovanniPasq/agentic-rag-for-dummies

    A modular Agentic RAG framework using LangGraph that implements hierarchical indexing, multi-agent parallel retrieval, and human-in-the-loop query clarification.

  • Project
    hhy-huang/HiRAG

    HiRAG is a Retrieval-Augmented Generation framework that uses hierarchical knowledge to improve the quality and comprehensiveness of LLM responses compared to standard RAG.