Grounded Extraction

This example implements Model as Sensor (1). A model proposes a customer name and quote; the host measures their occurrence in the source; rules decide whether to accept the claim. It makes the request, receipt and accept moves visible.

The program

Download grounded-extraction.fx, or copy this complete file:

world open.

doc1:Document[text -> "Acme AG (the Customer)."].

external readCustomer(?Document) using _llm[
    prompt -> "Extract the customer. Return JSON with rows of slot, value and quote. Use slot customer, copy the value and quote exactly, and return no other claims.";
    document -> ?Text; groundIn -> ?Text]
    <- ?Document:Document[text -> ?Text].

// Only the customer field is admitted, and both host checks must hold.
?Document[accepted(?Row) -> ?Value] <- ?Document:Document
    AND ?Call[tool#name -> "_llm"; tool#arg1 -> ?Document;
              tool#completed -> true; tool#status -> "succeeded";
              row(?Row, "slot") -> "customer";
              row(?Row, "value") -> ?Value;
              row(?Row, "quoteFound") -> true;
              row(?Row, "valueQuoted") -> true].

?Document[customer -> ?Value] <- ?Document:Document
    AND ?Document[accepted(?Row) -> ?Value].

?Document[readError -> ?Message] <- ?Document:Document
    AND ?Call[tool#name -> "_llm"; tool#arg1 -> ?Document;
              tool#completed -> true; tool#status -> "failed";
              error -> ?Message].

?- doc1[customer -> ?Customer].
?- doc1[readError -> ?Message].

The Document class condition restricts the subject. The receipt joins back to it through tool#arg1. Acceptance requires a successful completed call, the allowed slot customer, and both quoteFound and valueQuoted.

This is a small lexical grounding example. It does not prove that a source is authoritative or that a copied value is semantically the customer. A real application needs source permissions and suitable domain checks. Model confidence is not an acceptance condition.

Test without a model

The following authored receipts simulate the host’s output. They test rule behavior; they do not measure a model or independently test quote matching. Save the program and fixtures together, then run each case separately:

framex test grounded-extraction.fx grounded-extraction.accepted.test.fx
framex test grounded-extraction.fx grounded-extraction.rejected.test.fx
framex test grounded-extraction.fx grounded-extraction.failed.test.fx

Do not concatenate conflicting scenario receipts. Under this open-world program, a refused customer claim stays unknown; refusal does not establish the opposite claim.

Run against a model

After installation and model configuration:

python -m framex.tools grounded-extraction.fx --receipts-out extraction.receipts.fx
framex run grounded-extraction.fx extraction.receipts.fx
framex explain grounded-extraction.fx extraction.receipts.fx 'doc1[customer -> "Acme AG"]'

The first command runs the host. The second replays receipt facts offline; the third inspects the conclusion if the model supplied accepted evidence. Live output depends on the model and is not guaranteed to match the authored success case.

The host’s groundIn checks compute the two Boolean fields; a model cannot authorize its own claim by writing quoteFound: true. A success event still does not bypass the acceptance rule. If no acceptable value is returned, the application should derive a bounded repair or a question for a person.