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 persistence in traditional AI search agents. It aims to create a search experience similar to Perplexity, where the system becomes "stronger the more it is used" by implementing a structured, multi-agent pipeline and a hierarchical memory system that avoids the pitfalls of purely research-oriented, high-latency self-evolving agents.

How it works

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

  1. Routing: A router determines if the query can be handled by a specialized tool (e.g., weather, GitHub, arXiv) or needs the retrieval pipeline.
  2. Query Rewriting: The query is rewritten to include time and location context before retrieval.
  3. Hierarchical RAG (L1–L5): The system retrieves information across five layers in parallel:
    • L1 (QA Cache): High-frequency Q&A for millisecond responses.
    • L2 (Commonsense): Wikipedia-based vector index for general knowledge.
    • L3 (History): Incremental FAISS index of successful previous answers to capture user preferences.
    • L4 (Web): Real-time web search via providers like Tavily or Serper.
    • L5 (Knowledge Graph): Wikidata-based structured facts for multi-hop queries.
  4. Fusion & Calibration: Results are merged using Reciprocal Rank Fusion (RRF) and calibrated via Platt scaling to a unified confidence score. It uses a "quota-based" fusion to ensure multiple entities in a single query are all represented.
  5. Summary & Attribution: An LLM generates the final answer with citations, which are then validated to prevent "citation hallucinations."

Who it’s for

  • AI Search Beginners: Those looking for a stable, well-documented framework to start with.
  • AI Engineers: Developers wanting to implement a production-ready search agent with latency governance and safety guards.
  • AI Search Architects: Professionals interested in hierarchical memory RAG and entity disambiguation in knowledge graphs.

Highlights

  • 5-Layer Memory Stack: Combines cache, Wikipedia, user history, live web, and knowledge graphs for comprehensive retrieval.
  • Confidence Calibration: Maps diverse scoring metrics (cosine, rank, hybrid) to a unified probability of relevance.
  • Latency Governance: Implements layer-specific budgets and "soft timeouts" to ensure the front-end response time has an upper bound.
  • Prompt Injection Protection: A three-layer defense (cleaning, XML tagging, and system prompt guards) to secure untrusted web content.
  • Entity Disambiguation: Uses a combination of popularity (in-degree) and context signals to distinguish between entities with the same name.
  • Self-Improving: Asynchronously archives successful answers into L1 and L3 to speed up future identical or similar queries.

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
    infiniflow/ragflow

    An open-source RAG engine that combines deep document understanding and agentic retrieval to create high-precision context layers for LLMs using complex unstructured data.

  • Project
    SaiAkhil066/CORTEX-AI-SUPER-RAG

    Cortex AI SUPER RAG is a local, open‑source RAG app that lets you upload PDFs/Word docs, ask natural‑language questions, and receive answers with precise citations. It combines BM25 keyword search, FAISS vector search, optional reranking, entity‑graph expansion, and several LLM‑based tricks—all runnable via Ollama without cloud APIs. The repo includes a Streamlit UI, a full evaluation harness (e.g., FinanceBench benchmark), and Docker support.

  • Project
    java-up-up/nexus-agent

    An enterprise-grade AI agent platform that implements a complete loop from document governance and hybrid RAG retrieval to ReAct-based agentic reasoning and MCP tool integration.