The framework-free backbone for backend RAG.
Deterministic dual-channel retrieval and agent orchestration, with anti-fabrication and source provenance built in — no Dify, no LangGraph, no DSL. Just composable Python.
ACME_CN FY2024 REVENUE 为 1320 USD_M（来源：ACME_FY2024_Review.pptx · slide=2,table=1,row=REVENUE,col=FY2024）
查不到：REVENUE / ACME_CN / 2025 …未在事实表中找到。为避免误导，不提供任何推测数字。
When the data isn't there, the orchestrator deterministically refuses — never guesses.
Infrastructure decisions, made for you — and enforced in code.
No framework lock-in
Pure Python; bring your own everything. Drop into any backend.
Dual-channel
Deterministic numbers + narrative RAG, unified by an agent router.
Anti-fabrication + provenance
Enforced invariants, not prompt suggestions.
Office-document extraction
xlsx / pptx / pdf → structured facts, style- & color-aware.
Hybrid retrieval
CJK-aware BM25 + injectable vector + RRF fusion + optional LLM rerank.
FAQ short-circuit
SME-vetted answers bypass the LLM, behind conservative exclusion guards.
Built-in evaluation
Four-gate metrics + baseline regression gating.
Privacy-aware observability
Traces carry codes / counts / timings only.
Two channels, one router.
Exact numbers come from a deterministic structured channel; everything else flows through narrative RAG. A single agent router decides, then attaches provenance to the answer.
Exact facts from a deterministic fact table — never inferred.
Hybrid retrieval over prose, fused and optionally reranked.