Ora — Architecture

alpha

LLM-agnostic · any database · 3 lines of code

v2.0 Docs GitHub

↑ the product ↑

ora-app · workspace · localhost:8080

$ pip install ora-sql Copy

⭐ GitHub 🤗 HuggingFace 📖 Docs 💬 Discord

Ask anything about your data...

→

Revenue by region this quarter Average order value over time Monthly active users trend Inventory turnover by category

3-line start multi-source

import ora

db = ora.connect("postgresql://localhost/mydb")

result = db.query("Show me top 10 customers by revenue")

✓ DataFrame · 1.2s · $0.003 · BIRD 65%+ Copy

↓ speak to your data ↓

Your workspaces

Loading…

◉

Data Layer

connected sources · any dialect · SQLAlchemy 2.0

PostgreSQL

Snowflake

BigQuery

DuckDB

MySQL

Redshift

CSV / XLSX

Query Received

User's natural language query arrives — Ora orchestrates the full pipeline: anaphora resolution, intent classification, entity alias mapping, data coverage analysis, and cross-source planning before routing to agents.

query received

◈

Ora

query orchestrator & reasoning agent

↑ click any node to see live output

STAGE 1

Intent Analysis

anaphora · scope · ambiguity

STAGE 2

Entity Resolution

fuzzy match · alias mapping

STAGE 3

Data Coverage

source scan · gap detection

OUTPUT

Execution Plan

route · decompose · schedule

orchestratorroutes queries across agents
owns the full pipeline LangGraphStateGraph orchestration
conditional routing edges LiteLLMall LLM providers via one adapter
GPT-4o · Claude · Ollama Qdrantsemantic cache cosine ≥ 0.85
skips LLM on repeat queries SemanticMemorySQLite entity alias store
learns from every query

Schema Analysis

SchemaAgent introspects the DB → builds KnowledgeGraph → detects FK topology → serializes to M-Schema for token-efficient prompting.

schema analysis

◈

Schema Agent

knowledge construction · graph topology · vector indexing

↑ click any node to see live output

STAGE 1

DB Introspect

DDL · FK · cardinality

STAGE 2

Knowledge Graph

nodes · edges · layers

STAGE 3

FK Topology

Merkle · 3-pass detect

STAGE 4

Vector Index

M-Schema · NL→SQL

LangGraph5-node agent StateGraph
SchemaAgent class SQLAlchemy 2.0async dialect reflection
PostgreSQL · Snowflake · BQ SQLGlotAST parsing · transpilation
cross-dialect normalization Qdrantin-process vector store
NL→SQL example retrieval fastembedBAAI/bge-small-en-v1.5
zero-API-key embeddings DuckDBin-memory cross-source JOIN
data never leaves environment

Query Orchestration

LangGraph StateGraph routes NL query → CHESS schema pruning → parallel SQL generation → ReFoRCE 3-stage self-correction → result.

query orchestration

⬡

AgentSQL Orchestration

LangGraph pipeline · self-correcting · CHESS + ReFoRCE

STAGE 1

Schema Linker

CHESS LSH · RAG

STAGE 2

SQL Generator

3 candidates · parallel

STAGE 3

Self-Corrector

ReFoRCE · 3-stage loop

OUTPUT

Result

DataFrame · NL · Chart

LangGraphStateGraph per query
conditional edges LangChainTool adapters +
few-shot prompt chains LiteLLMUnified router — OpenAI
Anthropic · Bedrock · Ollama OpenTelemetryOne span per node
Langfuse + Prometheus FastAPISSE /query/stream
live trace events to UI

Specialist Agents

Decompose cross-source queries → fan out to each DB → DuckDB in-memory synthesis → NL response generation → LearningLoop trains on feedback.

specialist agents

◇

Specialist Agents

decompose · synthesize · respond · learn

AGENT

DecomposeAgent

cross-source split · DAG planner

AGENT

SynthesisAgent

DuckDB JOIN · in-memory merge

AGENT

ResponseGenerator

NL summary · chart · follow-ups

AGENT

LearningLoop

feedback → training pairs

DuckDBin-memory cross-source JOIN
data never leaves environment LangGraphsupervisor orchestrator
QueryOrchestrator class LiteLLMNL summary generation
follow-up question generation QdrantNL→SQL training pair store
thumbs-up → auto-indexed