Query a database you've never seen,
in your own words

sans_schema — a Semantic Query Gateway

sans-schema-playground.vercel.app

Open it now — you'll use it during the talk.

Every client hardcodes someone else's schema

• Your code says author_name. Theirs says writer. The next API says author.full_name.
• Every consumer re-learns every backend's vocabulary.
• Every rename breaks every consumer.
• And now AI-built clients guess field names — the seam should absorb a good-enough guess.

{want, where} — your words, their data

POST /query
{
  "want": ["book name", "writer", "cost"],
  "where": "science fiction under 25 dollars"
}

{
  "rows": [
    { "book name": "Vingt mille lieues sous les mers",
      "writer": "Jules Verne",
      "cost": 15.5 } ],
  "interpreted": {
    "want": { "writer": {
      "field": "books_view.author_name",
      "confidence": 0.95 } },
    "where": { "ast": { "op": "and", … },
               "confidence": 0.9 } }
}

want: fields in your vocabulary · where: a plain-language filter · the answer comes back in your keys, with an inspectable interpreted echo.

Natural-language easy on the outside, a deterministic contract on the inside — stable enough for a traditional application to build against.

The gateway never sees a schema config

• At startup it asks the database — column names, types, the DBA's column comments, a few sample values.
• That introspected schema becomes the resolver's system prompt — the vocabulary the model matches your want + where against.
• Point it at a different Postgres → nothing to rewrite. New columns show up on restart.

The schema block sits behind a prompt-cache boundary — after the first request it costs ~0.1× tokens per query.

How it stays safe

NL → validated predicate AST → execute. Never NL → SQL.

The shape is an hourglass: many protocols on top, many data sources at the bottom — one narrow waist everything must squeeze through.

A hijacked model can at worst mis-resolve inside the allowed schema — it can never emit SQL. You'll see the proof live: the playground echoes the exact SQL it ran, values as bound parameters.

Spike before build

The one risky bet — can a model resolve fields + compile filters reliably? — was measured before any gateway code existed.

• Production model (Gemini flash-lite) — field resolution (want): 100% · filter compilation (where): 98%
• Across 9 models, 3 vendors, on the same 52-case adversarial exam (4 schemas, 125 fields, 40 filters): want 99–100% · where 80–100% — the spread is why the model sits behind a swappable interface.
• scored by execution equivalence (same rows selected)
• Below-confidence requests are refused, not guessed — the gate is a feature.

The spike harness is frozen and re-runnable — every resolver change re-takes the same exam.

Live demo

sans-schema-playground.vercel.app

Made-up field names → real answers. Every run shows its work: what the gateway understood (resolved fields, confidence, cache hits, the AST), the exact SQL with bound parameters, and a copy-paste curl of the whole request.

Try it with your own data

1. git clone + docker build
2. docker run with your DATABASE_URL + your LLM key
3. curl it in your own words

sans-schema-playground.vercel.app/own-data

Deliberately not solved yet

• Authorization — the gateway is a curated single view today; field-level authz is its own milestone.
• Messy schemas — accuracy is measured on a clean schema; cryptic legacy columns are the next benchmark.
• Agent-traffic cache economics — the cost model leans on cache hits; agents invent novel phrasings.

Stated limits are how you know the numbers are honest.

What's next — swap the hourglass ends

• New RequestAdapters — GraphQL / Protobuf requests in, the same resolver behind them.
• Richer IR — joined tables at the waist.
• New Connectors — data catalogs and GraphQL backends as sources.

Plus symbolic dates (bind_today) — compile “published over 20 years ago” date-independent, bind today at execute time, so the cache survives midnight.

Thanks

• Repo: github.com/SansWord/sans_schema
• Playground: sans-schema-playground.vercel.app
• Contact: sansword@gmail.com
• Portfolio: sansword.github.io/resume
• LinkedIn: linkedin.com/in/sansword