How p-breslin/zero-touch works
Project type and purpose
• This is a data-processing orchestration engine (not a typical web app). It uses the Agno Agent Framework to build a pipeline of AI “agents” that extract, plan, generate, and execute SQL against engineering data (GitHub/JIRA) and resolve identities.
• Its goal is to infer organizational structure automatically and (eventually) populate an Object-Oriented Graph in ArangoDB.
File structure
• Top-level “configs/” holds YAML configs for agents, prompts, and query templates.
• “services/” encapsulates DB connectivity, engine creation, and data export logic (Snowflake → CSV).
• “scripts/” contains setup scripts to provision DuckDB, ChromaDB, and ArangoDB.
• “models/” defines Pydantic schemas (SQLPlan, AggregatedData, IdentityList, etc.).
• “src/” houses the core logic:
• orchestrator.py (pipeline driver)
• agents/ (builder, base class)
• paths.py, test.py
• “dev/” and “utils/” contain helper scripts and logging setup.
Extracted insights from README
• Four main layers: Data Sources, Knowledge Base (ChromaDB), Agno Agent Framework, and (future) Graph DB (ArangoDB).
• Pipeline steps:
1. Orchestrator kicks off
2. KnowledgeBase Agent retrieves schema info via ChromaDB
3. Planner Agent creates an SQLPlan
4. SQL Constructor Agent builds SQL strings
5. SQL Executor Agents run queries in parallel on DuckDB
6. Orchestrator aggregates results
7. Identity Agent resolves unique persons
8. (Future) Graph Agent will load results into ArangoDB
a. Main components
• External Data Sources: GitHub, JIRA (Snowflake)
• Setup Scripts: SnowflakeDownload → DuckDBSetup → ChromaDBSetup → ArangoDBSetup
• Databases:
• DuckDB (local mirror)
• ChromaDB (vector store for schema KB)
• ArangoDB (OOG graph)
• Service Layer: connection_manager, create_engine, snowflake_download
• Agent Framework Layer:
• Orchestrator
• Agents (KnowledgeBase, Planner, SQLConstructor, SQLExecutor, Identity)
• Models: Pydantic schemas for data contracts
• Configs & Utils: YAML configs, helpers, logging
b. Relationships & interactions
• Data flow: GitHub/JIRA → SnowflakeDownload service → CSV files → DuckDB
• ChromaDBSetup loads schema JSON → ChromaDB
• Orchestrator reads configs → instantiates Agents
• KnowledgeBase Agent queries ChromaDB → returns schema info
• Planner Agent consumes schema info → outputs SQLPlan
• SQLConstructor Agent transforms SQLPlan → SQL strings
• SQLExecutor Agents execute SQL on DuckDB → return raw rows
• Orchestrator aggregates rows → AggregatedData
• Identity Agent resolves persons → IdentityList
• (Future) Graph Agent writes entities/relations to ArangoDB
c. Architectural patterns & principles
• Extreme Task Decomposition: very fine-grained agents each with a single responsibility
• Centralized orchestration: orchestrator directs sequence, parallelism, error checks
• Clear data contracts via Pydantic models
• Configuration-driven: YAML files decouple code from prompts and settings
d. Key technologies
• Python 3.12+, Agno Agent Framework, OpenAI/LLMs
• SQLAlchemy (DuckDB, Snowflake)
• ChromaDB (vector knowledge base)
• DuckDB (local dev mirror)
• ArangoDB (graph store)
• Pydantic models, python-dotenv, uv for builds