Self-Improving SQL Agent

LangGraph Text-to-SQL agent over a music-store schema that learns from its own failures.

It generates SQL, runs it in a SELECT-only sandbox, reflects on errors, and writes durable lessons into a playbook. Holdout evals compare three modes (A / B / C) so improvement is measurable — not anecdotal.

Optional Mem0 (or a local store) keeps user preferences. The execute path is also exposed as an MCP tool.

Repo: https://github.com/ssk525/self-improving-sql-agent

Why this exists

Most Text-to-SQL demos stop at “LLM wrote a query.” Production systems need:

1. Guardrails — never run arbitrary SQL

2. Recovery — retry with reflection when the executor fails

3. Memory — don’t re-learn the same join mistake every session

4. Measurement — frozen holdout EX%, not cherry-picked screenshots

This repo is scoped to those four.

Related MCP server: readonly-postgres-mcp

Architecture

question
  → schema linker
  → user memory (optional)
  → playbook retrieve
  → generate SQL
  → SELECT / WITH-only guardrail
  → sandbox execute
  → on error: reflect → update playbook → retry (max 3)
  → result table

Primary metric: execution accuracy (EX%) — predicted SQL result rows match gold SQL on a frozen 20-question holdout (column names ignored).

Eval results

From make eval with Ollama gemma4:e4b (see evals/results.md):

Takeaway: On this local model, retries rarely fire — misses are usually wrong-but-valid SQL. Playbook did not lift holdout EX% here (noise risk with small models). The point of the harness is to measure that, not to claim magic gains. Re-run with Groq/Gemini/gpt-4o-mini before quoting cloud numbers.

Quickstart

git clone https://github.com/ssk525/self-improving-sql-agent.git
cd self-improving-sql-agent
python3.12 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env
# set GROQ_API_KEY / GOOGLE_API_KEY / OPENAI_API_KEY, or USE_OLLAMA=1
make init
make test
make run                      # http://localhost:8501
make eval                     # writes evals/results.md

PostgreSQL (optional; needs Docker):

# DATABASE_URL=postgresql://sqlagent:sqlagent@127.0.0.1:5432/sqlagent
make up

Design decisions (interview talking points)

1. SELECT-only sandbox — Text-to-SQL without a guardrail is a drop-table demo waiting to happen.

2. Playbook ≠ chat history — durable, retrieved lessons across queries; Mode C isolates that effect on holdout.

3. EX% over string-match SQL — two different SQL strings can be equivalent; comparing result multisets is fairer.

4. Learn / holdout split — playbook may train on 30 learn questions; scoring uses 20 frozen holdout IDs only.

5. Ablations over vibes — A/B/C exists so you can say what helped, and admit when it didn’t.

6. MCP wrapper — same execute path callable from MCP clients for tool-use interviews.

Stack

Python 3.12 · LangGraph · PostgreSQL 16 / SQLite · Pydantic · Mem0 (optional) · MCP · Streamlit · pytest · Docker Compose

Resume one-liner

Built a LangGraph Text-to-SQL agent with SELECT-only execution, reflection retries, persistent playbook memory, and holdout execution-accuracy evals (75% EX% on 20 frozen questions with local gemma4:e4b; A/B/C ablations reported in-repo).

License

MIT