PRAGPengyi RAG Intelligence Group

RAG research and deployment platform

Retrieve.
Verify.
Deploy.

PRAG studies how knowledge systems behave, proves which architecture fits the task, and deploys only after human review.

Open the benchmark system →
An editorial RAG laboratory showing documents, retrieval paths, a knowledge graph, verification and benchmark charts
Research → Evaluation → Decision → Production
01 / Discover02 / Analyze03 / Benchmark04 / Audit05 / Deploy06 / Re-measure
01 / Research map

Architecture is a testable choice

Five RAG families.
No automatic winner.

01

Vanilla

Chunk, embed, retrieve, generate. The minimum credible RAG baseline.

Best forFocused document QA
02

Advanced

Query rewrite, hybrid search, reranking and multi-hop retrieval.

Best forHigher recall and precision
03

Graph

Entities and relationships support global synthesis and connected reasoning.

Best forFinancial and organizational networks
04

Agentic

A bounded planner chooses when, where and how often to retrieve.

Best forMulti-step research workflows
05

Multimodal

Retrieval across text, tables, figures, images, video and complex PDFs.

Best forMixed-format evidence
02 / Agent Team

Auditable division of responsibility

Agents propose.
Evidence and people decide.

01

Connector

Finds primary research, repositories, datasets and benchmark evidence.

02

Analyzer

Extracts assumptions, architecture, strengths, weaknesses and evidence quality.

03

Deployment

Maps requirements to candidate systems, budgets and expected trade-offs.

GATE

Human Audit

Confirms need, safety, cost, source rights and production boundaries.

04

Executor

Builds the approved system with tests, telemetry, rollback and version control.

05

Orchestrator

Routes the lifecycle, enforces schemas and preserves provenance.

03 / Benchmark OS

Compared with what?

A RAG system is only as good as its evidence.

L0No RAG
L1Vector
L2Hybrid + Rerank
L3Task-specific
L4Previous Best
04 / Deployment map

One platform, different requirements

Architecture follows the project.

PAAT

Quant research memory

Papers, factors, experiments and evidence relationships.

Candidate: Hybrid / Graph / Agentic
PDAT

Data documentation

Schema, lineage, dictionaries, ownership and quality controls.

Candidate: Hybrid / Graph
PIT

Interview knowledge

Questions, project evidence, company research and follow-up trees.

Candidate: Vector + Reranker
HARNESS

Workflow intelligence

Tools, runbooks, incidents, traces and operational memory.

Candidate: Agentic with strict bounds
Release contract Problem → Baseline → Experiment → Result → Trade-off → Human Audit → Production