01 / Agentic AI & LLMOps
In progressEnterprise Multi-Agent Knowledge & Research Assistant
Target role · GenAI / Agentic AI Engineer
A cyclical Planner, Retriever, and Fact-Checker workflow combining hybrid and vectorless RAG with enterprise guardrails.
- Target: 34% lower context hallucination
- Evaluate: faithfulness · relevancy · recall
Problem & direction
Core problem: Vanilla RAG can hallucinate on complex queries requiring cross-document synthesis, source verification, and structured reports.
An in-progress research system designed around a stateful team of agents. Its planned workflow combines hybrid retrieval, a structured vectorless fallback, fact checking, security middleware, and regression evaluation.
How it works
- 01
Plan and validate the query
- 02
Retrieve with Qdrant or Chroma plus BM25, with structured SQL or JSON fallback
- 03
Fact-check and synthesize across sources in LangGraph
- 04
Evaluate, trace, and stream the result through FastAPI SSE
Engineering details
Stateful agent orchestration
LangGraph models cyclic workflows so agents can retain state, reflect on intermediate results, and revisit an answer before returning it.
Grounded knowledge retrieval
Qdrant provides the vector retrieval layer for self-reflective RAG, connecting the answer-generation workflow to relevant source context.
Typed outputs and guardrails
Pydantic v2 enforces response schemas. PII redaction callbacks and prompt-injection defenses address sensitive data and untrusted input.
Evaluation plan
Ragas or DeepEval will measure faithfulness, relevancy, and context recall. The 34% hallucination reduction is a target until validated.