LARS - AI That Speaks SQL

Your team knows SQL. Why learn Python for AI?

Add AI operators directly to your SQL queries — from your existing SQL client, on your existing databases.

SELECT * FROM support_tickets
WHERE description MEANS 'urgent customer issue';

That's it. No notebooks. No orchestration code. No vector database to provision. Just SQL with semantic understanding (and a declarative workflow engine underneath). Express intent, not patterns — especially when you don't know what you're looking for.

• Use your existing SQL client: PostgreSQL wire protocol (lars serve sql)
• Your data never moves: DuckDB federation across Postgres/MySQL/BigQuery/Snowflake/S3/files
• Cached + cost-attributed: query LLM calls via all_data / sql_query_log
• Optional Studio UI: inspect runs, takes, costs, and "what the model saw" (not required)

One line. That's all it takes.

Before: Regex, LIKE patterns, and brittle keyword matching.

SELECT * FROM tickets
WHERE description LIKE '%urgent%'
   OR description LIKE '%critical%'
   OR description LIKE '%asap%'
   -- still misses "need this fixed immediately"

After: One line that understands meaning.

SELECT * FROM tickets
WHERE description MEANS 'urgent customer issue'

What Can You Do?

-- Filter by meaning, not keywords
SELECT * FROM products
WHERE description MEANS 'eco-friendly'

-- Score relevance (0.0 to 1.0)
SELECT title, description ABOUT 'sustainability' AS relevance
FROM reports
ORDER BY relevance DESC

-- Semantic deduplication
SELECT SEMANTIC DISTINCT company_name FROM leads

-- Find contradictions (compliance, fact-checking)
SELECT * FROM disclosures
WHERE statement CONTRADICTS 'no material changes'

-- Summarize groups
SELECT category, SUMMARIZE(reviews) AS summary
FROM feedback
GROUP BY category

-- Group by auto-discovered topics
SELECT TOPICS(title, 5) AS topic, COUNT(*) AS count
FROM articles
GROUP BY topic

-- Vector similarity search
SELECT * FROM docs
WHERE title SIMILAR_TO 'quarterly earnings report'
LIMIT 10

-- Ask arbitrary questions
SELECT
  product_name,
  ASK('Is this suitable for children? yes/no', description) AS kid_friendly
FROM products

-- plus lots more...

100+ built-in operators for filtering, logic, transformation, aggregation, data quality, parsing, and more.

Quick Start

# install
pip install larsql

# set your LLM API key (OpenRouter, or see docs site for others)
export OPENROUTER_API_KEY=sk-or-v1-...

# set up clickhouse (docker or existing DB)
docker run -d \
  --name lars-clickhouse \
  --ulimit nofile=262144:262144 \
  -p 8123:8123 \
  -p 9000:9000 \
  -p 9009:9009 \
  -v clickhouse-data:/var/lib/clickhouse \
  -v clickhouse-logs:/var/log/clickhouse-server \
  -e CLICKHOUSE_USER=lars \
  -e CLICKHOUSE_PASSWORD=lars \
  clickhouse/clickhouse-server:25.11

# create & populate a project directory for the starter files
lars init my_lars_project ; cd my_lars_project

# init the database and refresh the metadata
lars db init 

# start the SQL server (PostgreSQL wire protocol)
lars serve sql --port 15432

# connect with any SQL client (default is lars/lars - proper auth coming soon)
psql postgresql://localhost:15432/default

# optional - start the web UI admin / studio tool
lars serve studio
# runs at http://localhost:5050

That's it. Run semantic queries from DBeaver, DataGrip, psql, Tableau, or any PostgreSQL client. For a full end-to-end setup (ClickHouse + sample data + Studio UI), see the Quickstart Guide.

How It Works

LARS uses query rewriting - your semantic SQL is transformed into standard SQL with UDF calls that execute LLM operations. Your database stays untouched.

WHERE description MEANS 'urgent'
         ↓
WHERE semantic_matches('urgent', description)
         ↓
UDF runs LLM → returns true/false

Results are cached - same query on same data costs zero after the first run.

Every semantic UDF call is also logged (model, tokens, cost, duration) into queryable "magic tables":

SELECT session_id, cell_name, model, cost, duration_ms
FROM all_data
WHERE is_sql_udf = true
ORDER BY timestamp DESC
LIMIT 20;

Every semantic operator is backed by a cascade file under cascades/semantic_sql/ - edit YAML to change behavior or create your own operator. If you want a visual view of the same execution data, Studio is a UI over these logs (optional).

Wait, it gets weirder.

Semantic SQL is just the beginning. Under the hood, LARS is a declarative agent framework for building sophisticated LLM workflows.

The Problem It Solves

Every LLM project eventually becomes this:

for attempt in range(max_retries):
    try:
        result = llm.call(prompt)
        if validate(result):
            return result
        prompt += f"\nError: {validation.error}. Try again."
    except JSONDecodeError as e:
        prompt += f"\nFailed to parse: {e}"
# 47 lines later... still doesn't work reliably

The LARS Solution

Run multiple attempts in parallel. Filter errors naturally. Pick the best.

- name: generate_analysis
  instructions: "Analyze the sales data..."
  takes:
    factor: 3  # Run 3 times in parallel
    evaluator_instructions: "Pick the most thorough analysis"

Instead of serial retries hoping one succeeds, run N attempts simultaneously and select the winner. Same cost, faster execution, higher quality output.

Declarative Workflows (Cascades)

Define multi-step agent workflows in YAML:

cascade_id: analyze_data
cells:
  - name: query_data
    tool: sql_data
    tool_inputs:
      query: "SELECT * FROM sales WHERE date > '2024-01-01'"

  - name: analyze
    instructions: |
      Analyze this sales data: {{ outputs.query_data }}
      Create visualizations and summarize key trends.
    skills:
      - create_chart
      - smart_sql_run
    takes:
      factor: 3
      evaluator_instructions: "Pick the most insightful analysis"
    handoffs: [review]

  - name: review
    instructions: "Summarize the findings"
    context:
      from: [analyze]

Key Concepts

Database Support

LARS connects to your existing databases:

• DuckDB (default, in-memory or file)
• PostgreSQL, MySQL, ClickHouse
• BigQuery, Snowflake
• S3, Azure, GCS (Parquet, CSV, JSON)

Your data stays where it is. LARS queries it federated-style. Join across DB boundaries.

LLM Providers

Works with any LLM via LiteLLM:

• OpenRouter (default) - access to 200+ models, excellently granular cost tracking
• OpenAI, Anthropic, Google
• Ollama (local & remote models, zero cost)
• Azure OpenAI, AWS Bedrock, Vertex AI

Installation Options

# Basic
pip install larsql

# With browser automation (Playwright)
pip install larsql[browser]

# With local models (HuggingFace)
pip install larsql[local-models]

# Everything
pip install larsql[all]

Running Cascades

# Run a workflow
lars run cascades/example.yaml --input '{"task": "analyze sales data"}'

# With model override
lars run cascades/example.yaml --model "anthropic/claude-sonnet-4"

Studio Web UI (Optional)

# Launch the visual interface
lars serve studio

# Access at http://localhost:5050
# - SQL IDE with semantic operators
# - Cascade runner (incl. takes + winners)
# - Context inspector ("what the model saw")
# - Cost explorer (by query/cascade/model)

Documentation

Full documentation at larsql.com

• Docs hub - Full reference

• Quickstart Guide - Get running in 10 minutes

• Studio Web UI - Optional UI for debugging cost/context/takes

• Semantic SQL - Query rewriting, caching, annotations, observability

• Built-in Operators - All 100+ operators

• Vector Search & Embedding - SIMILAR_TO, LARS EMBED, hybrid search

• Cascade DSL - Workflow configuration

• Takes & Evaluation - Parallel execution patterns

• SQL Connections - Connect 18+ data sources via DuckDB

• AI Providers - OpenRouter, Vertex AI, Bedrock, Azure, Ollama

• Tools Reference - Available skills & integrations

Example: Create Your Own Operator

Any cascade can become a SQL operator. No Python required.

# cascades/semantic_sql/sentiment_score.cascade.yaml
cascade_id: sentiment_score
sql_function:
  name: SENTIMENT_SCORE
  operators:
    - "SENTIMENT_SCORE({{ text }})"
  returns: DOUBLE
  shape: SCALAR
cells:
  - name: score
    model: google/gemini-2.5-flash-lite
    instructions: |
      Rate the sentiment of this text from -1.0 to 1.0.
      TEXT: {{ input.text }}
      Return only the number.

Worried about the output? Me too. Run validations or multiple takes (on multiple models), all within a SQL call.

Now use it:

SELECT product_id, AVG(SENTIMENT_SCORE(review)) AS sentiment
FROM reviews
GROUP BY product_id
HAVING sentiment < -0.3

Contributing

Issues welcome at github.com/ryrobes/larsql

License

O'SASSY License (basically MIT)