l0l1.skelfresearch

How l0l1 works

l0l1 is a deterministic-plus-LLM layer that sits between whoever wrote the SQL and the warehouse that will run it. It reasons about four boundaries and applies protection at each one. 

   author (human · ChatGPT · internal tool)
              │  ① user → tool   ── PII scan (Presidio + regex)
              ▼
        ┌───────────────────────────────┐
        │            l0l1               │
        │  scan · validate · learn      │
        └───────────────────────────────┘
          │                       │
  ② tool → provider        ③ tool → warehouse
  anonymize before send    respect granted scope
          │                       │
          ▼                       ▼
   OpenAI / Anthropic      PostgreSQL · MySQL
   (validation reasoning)  SQLite · DuckDB
                                  │
                          ④ warehouse → output
                          annotate PII in results
                                  ▼
                             result rows

The four boundaries

1

User → tool

Whatever a person or an LLM produced arrives at l0l1: a prompt, a query, or both.

l0l1 · PII detection runs here. Presidio plus regex scan prompts and query text for SSNs, emails, phone numbers, and card numbers.

2

Tool → model provider

If validation uses OpenAI or Anthropic, query text and schema context leave the local process.

l0l1 · Detected PII literals are anonymized before this crossing, so raw identifiers are not forwarded to the provider.

3

Tool → warehouse

The query runs against the actual database — PostgreSQL, MySQL, SQLite, or DuckDB.

l0l1 · l0l1 respects whatever access scope you grant. It reads schema to validate; it does not widen your permissions.

4

Warehouse → output

Result rows come back and may be stored, cached, or surfaced to a user.

l0l1 · PII columns can be annotated in output, and successful queries are sanitized of literals before being learned as patterns.

The validation pipeline

  1. 1

    Ingest

    A query (and optional prompt) enters via CLI, REST API, Jupyter magic, or the VS Code/LSP editor integration.

  2. 2

    Scan

    services/pii_detector.py runs Presidio + regex over the text and anonymizes detected literals.

  3. 3

    Introspect

    services/schema_service.py reads and caches the live schema of the connected database.

  4. 4

    Validate

    The query is checked against real columns and relationships; an LLM reviewer (OpenAI/Anthropic) reasons about intent.

  5. 5

    Learn

    services/learning_service.py sanitizes the approved query into a shape pattern and stores it per workspace.

  6. 6

    Suggest

    Later, stored patterns are surfaced as completions for the next person writing similar SQL.

All features Glossary Quickstart →