Skip to content

Latest commit

 

History

14 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Atlas — Agentic Analytics Assistant

CI Python 3.11 License: MIT

A production multi-agent system that answers analytical questions over data. A planner decomposes the question; tool-using sub-agents run SQL/analytics, hybrid retrieval (BM25 + dense vectors with reciprocal-rank fusion), and sandboxed code execution; a synthesizer composes a cited answer. Built with the production concerns that demos skip: observability/tracing, evaluation, caching, and streaming.

Why it exists: most "agent" repos are a single prompt in a while loop. Atlas is an agent system — planning, typed tools, hybrid retrieval, tracing, and evals you can actually trust.

Architecture

flowchart TB
    Q[User question] --> PL[Planner agent<br/>decompose + route]
    PL --> ORCH{Orchestrator}
    ORCH --> SQL[SQL / analytics tool]
    ORCH --> RET[Hybrid retrieval tool<br/>BM25 + dense + RRF + rerank]
    ORCH --> CODE[Sandboxed code tool]
    SQL & RET & CODE --> SYN[Synthesizer<br/>answer + citations]
    SYN --> OUT[Streamed answer]
    ORCH -. spans .-> TR[(Tracing / observability)]
    SYN -. scored by .-> EV[(Evals — uses lens-llm-eval)]
Loading

What makes it different

  • Planner → tools → synthesis, not a single mega-prompt.
  • Hybrid retrieval — lexical (BM25) + dense vectors fused with reciprocal rank fusion, then reranked. Beats either retriever alone.
  • Typed tool registry — tools declare typed schemas; the orchestrator validates calls.
  • Observability first — every step emits a span; runs are replayable.
  • Evaluated — answer quality gated by lens-llm-eval, the companion eval framework.

Tech stack

Python 3.11 · multi-agent orchestration (LangGraph-style) · BM25 + dense embeddings + FAISS · reranking · Anthropic (Claude) + open models · FastAPI (streaming) · tracing/observability · Docker

Status

🚧 Built in public, in phases — see ROADMAP.md. Each phase ships tested code + a design note in docs/.

Quickstart

pip install -e ".[dev]"
pytest

License

MIT © Venkateswarlu Nagineni

About

Agentic analytics assistant: planner + typed tools + hybrid retrieval (BM25+dense+RRF), with tracing, evals, and streaming

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages