Fact Extractor

View as Markdown

A RocketRide text-processing node that extracts configured records from documents or tables, with optional LLM validation and source provenance. Choose it over extract_data when each emitted value needs a traceable source location or a second extraction check.

What it does

The node buffers document and table context for an input object, then extracts one JSON record per logical record when that object closes. Its prompt includes the configured target fields and requires a _provenance object for every record. By default it sends the results through a second LLM prompt that rechecks the cited source, corrects or drops unsupported records, and attaches validation information. It emits the reconciled list as an answer and/or as one document per record. Use extract_data when a consolidated table is sufficient and there is no need to preserve per-record provenance or run a validator pass.

Connections

Connection	Required	Description
llm	yes	LLM used for fact extraction and, when enabled, validation.

Lanes

Lane in	Lane out	Description
table	answers	Emits the reconciled fact list as one JSON answer.
table	documents	Emits one JSON document for each reconciled fact.
documents	answers	Extracts from incoming documents and emits the reconciled fact list as one JSON answer.
documents	documents	Extracts from incoming documents and emits one JSON document for each reconciled fact.

Configuration

Configure the target fields, then decide whether the second validation pass and provenance should be included. The node has one hidden default profile, so these choices apply directly to its extraction behavior.

Fields

Provide between one and 32 target fields. Each field has a name, a type selected from the declared values, and an optional default value; the node places these in the LLM prompt. The extractor is instructed to infer a value from context when an exact field label is absent, so use names that express the facts you need rather than merely copying source headers. A field with a blank name or type is skipped at startup with a warning and is not available to the extractor or validator.

Field types and defaults

The type tells the LLM how to interpret the target value. Use text for unconstrained values, numeric and date/time types for values that should be recognized in those forms, and specialized types such as email, url, phone, or json when their structure matters. The default value is prompt context, not a post-processing guarantee, so select one only when it is a sensible fallback for a missing fact. Keep types and defaults consistent with the source you expect to validate; the validator checks the extraction result against the same configured fields.

Validate

Validation is on by default. With it enabled, the node runs a second prompt on the same LLM connection against the candidate records and source context. The validator can correct a value, repair its cited table location, lower its confidence, or remove a record that the source does not support. Turn it off when the extra prompt cost and latency are not worth the second check; in that mode the first-pass records are emitted without _validation annotations. If the validator returns no usable records, the node retains the first-pass candidates instead of discarding them.

Include Provenance

This setting defaults to on. The extraction prompt asks for page, table_id, row, col, source_text, and confidence in _provenance; table records use the node's [TABLE <id>] fences and document inputs can carry [Page N] markers. Turn it off only when downstream consumers do not need source traceability, because the node removes _provenance from the final records after validation. It does not disable validation itself.

Notes

Buffered input and output

The node consumes incoming content with preventDefault and emits derived facts only on close, so downstream lanes never see its original input or partial LLM output. Table input is fenced with sequential table IDs. On the documents lane, documents marked metadata.isTable are sent to the table buffer; other content is buffered as free text, and a page metadata value is preserved in a page marker for the LLM. Document output serializes each fact as JSON; its isTable metadata is true when the fact's provenance has a table ID.

Record shape

The built-in prompts request one object for each logical record, not a separate object for every target field. For table data, the extractor is told to treat the first row as a header and use a zero-based data-row index in provenance. The LLM supplies provenance on a best-effort basis, so it may return null for a source location that the provided context cannot establish.

Schema

Field	Type	Description	Default
facts.column	string	Field Name of the target fact field	"field"
facts.defval	string	Default Value	""
facts.fields	array
facts.include_provenance	boolean	Include Provenance Auto-attach provenance (page, table id, row, column, source text, confidence) to every emitted fact.	true
facts.profile	string		"default"
facts.type	string	Type	"text"
facts.validate	boolean	Validate Run a second validator prompt pass that double-checks extracted facts and reconciles on disagreement.	true