schemalytics 1.0.1

DB schema extraction to semantic layer for self-service analytics with local LLMs

Schemalytics

Semantic Layer for Self-Service Analytics

Schemalytics transforms your database into an LLM-ready semantic layer that enables self-service analytics. It analyzes your schema, generates dimensional models, and creates comprehensive metadata that LLMs can use to write accurate SQL queries—all running locally with complete privacy.

The tool generates a complete dbt project as the implementation layer, following dimensional modeling best practices with medallion architecture (Bronze → Silver → Gold).

Key features:

• Semantic layer generation - LLM-ready metadata with metrics, relationships, and query patterns
• Self-service analytics - Enable natural language queries against your data
• Privacy-first - Runs on local LLMs (Ollama) by default; Anthropic supported via env var
• Agentic pipeline - Five focused AI agents infer industry, metrics, and table roles from schema metadata alone
• Interactive refinement - Review and refine the generated data model through natural language feedback
• Fine-tuned modeling agents - Agents 3, 4a, and 4b have dedicated QLoRA fine-tuned Qwen3.5-4B models trained on real production schemas

Quick Start

1. Install prerequisites

# Install Ollama (default provider)
curl -fsSL https://ollama.com/install.sh | sh
ollama pull gemma3:4b                               # default model (Agents 1, 2, 5)

# Fine-tuned models for Agents 3, 4a, 4b (used by default)
ollama pull nichr0/schemalytics-classification-agent
ollama pull nichr0/schemalytics-silver-agent
ollama pull nichr0/schemalytics-gold-agent

# Install Schemalytics
pip install schemalytics

2. Generate semantic layer + dbt project

schemalytics generate \
  -c postgresql://user:password@localhost/mydb \
  -o ./dbt_project

3. Agentic pipeline + interactive refinement

• Agent 1 infers your industry and domain from schema metadata
• Agent 2 suggests metrics, goals, and reporting grain
• Agent 3 classifies each table as fact, dimension, bridge, or reference
• You review a consolidated summary and correct anything wrong
• Agent 4 generates a full modeling plan; you refine it with natural language ("make revenue weekly", "add customer LTV")
• Press Enter to approve and generate the dbt project

Optional: use Anthropic instead of Ollama

SCHEMALYTICS_LLM_PROVIDER=anthropic \
ANTHROPIC_API_KEY=sk-ant-... \
schemalytics generate -c postgresql://localhost/mydb -o ./dbt_project

Fine-Tuned Models

Three Qwen3.5-4B models are trained on real production schemas and used by default for their respective agents:

Model	Agent	Purpose	Default?
nichr0/schemalytics-classification-agent	Agent 3	Table classification (fact/dim/bridge/reference)	Yes
nichr0/schemalytics-silver-agent	Agent 4a	Silver layer plan (dim_*, fct_*)	Yes
nichr0/schemalytics-gold-agent	Agent 4b	Gold layer plan (agg_*)	Yes

All models: unsloth/Qwen3.5-4B base · QLoRA · Q4_K_M quantized · ~2.6 GB each

Agents 1, 2, and 5 use the general Ollama model (gemma3:4b by default, overridable via SCHEMALYTICS_OLLAMA_MODEL).

Per-agent model override:

# Override a specific agent's model
SCHEMALYTICS_AGENT3_MODEL=nichr0/schemalytics-classification-agent \
SCHEMALYTICS_AGENT4A_MODEL=nichr0/schemalytics-silver-agent \
SCHEMALYTICS_AGENT4B_MODEL=nichr0/schemalytics-gold-agent \
schemalytics generate -c postgresql://... -o ./dbt_project

Attribution — all models are built on Qwen3.5 by Alibaba Cloud (Qwen License).

What You Get

• Semantic layer (semantic_layer.yml) - Complete metadata for LLM-powered analytics
• Bronze models - Raw data staging layer (stg_<schema>_<table>)
• Silver models - Facts (fct_*) and dimensions (dim_*) in star schema
• Gold models - Pre-aggregated metrics (agg_<grain>_<metric>)
• Documentation - Auto-generated schema.yml files

CLI

# Full agentic pipeline
schemalytics generate -c postgresql://user:pass@localhost/db -o ./dbt_project

# Schema extraction only
schemalytics extract -c postgresql://user:pass@localhost/db -o schema.json

License

MPL 2.0 • Built by NiChr0

Fine-tuned models are based on Qwen3.5 (Qwen License) by Alibaba Cloud.