Everything K-culture — comebacks to K-beauty, straight to your inboxGet it in your inbox

METAL MEDIA

Irenezhangtt/AskingMe-Agent

17Python

An AI assistant for company policy questions that only searches harder when it's unsure

AskingMe Agent answers employee questions about internal company policies like HR, expenses, and system access. Instead of always running a heavy search process, it only rewrites and re-ranks queries when the initial search results look weak, balancing speed and accuracy. Separate specialist agents handle HR, expenses, access, and general questions, running in parallel for questions that span multiple domains.

What it does

  1. Built with a FastAPI backend, React frontend, Anthropic's Claude models, the ChromaDB vector database, and Redis for memory.
  2. Uses adaptive retrieval: it tries direct vector search first, and only rewrites the query and reranks results with an LLM when the retrieval confidence is low.
  3. Routes questions to four specialist agents (Expense, HR, Access, General); compound questions across domains trigger multiple specialists running concurrently.
  4. Manages the full lifecycle of policy documents (PDF, DOCX, TXT, Markdown, JSON) including duplicate detection, approval, versioning, and archival.
  5. Its internal evaluation reports grounded answer accuracy rising from 61.5% (LLM only) to 86.5% (full pipeline), and hallucination rate dropping from 24.0% to 5.0% over the same comparison.

Why it matters

For organizations dealing with scattered, frequently updated internal policies, this shows a practical pattern for keeping heavier retrieval steps on standby instead of running them every time, which helps control both latency and cost. It also builds in a safeguard that routes sensitive or exceptional cases to human review rather than letting the AI decide, which matters for real deployment.

Terms in this repo

  • RAG · a method that retrieves relevant documents before answering, using them as grounding evidence
  • ChromaDB · a vector database used to search documents by semantic similarity
  • embedding · a numeric vector representation of text used to measure meaning similarity
  • hallucination · when an AI generates a plausible-sounding but false or ungrounded answer
  • reranking · reordering an initial set of search results to improve accuracy

Repository description (English)

LLM-powered enterprise AI agent with adaptive RAG and multi-agent orchestration for company internal policy Q&A.

Open on GitHub

Trending repos

All repos →

Latest from METAL MEDIA