How it works
A compiler for unstructured text
C2F is not RAG. Not a knowledge graph. Not prompt engineering. It's a structured compilation pipeline that produces governed data products.
The pipeline
1
Grounding
Extract evidence spans from source text
2
Normalization
Types, units, negation, temporal refs
3
Schema Discovery
MECE taxonomy of fields
4
Question Gen
Typed slots from leaf nodes
5
Filling
Extract + version per document
6
Evaluation
Coverage, agreement, evidence rate
Grounding-first
- No value without evidence. Every extraction is anchored to source text.
- Uncertainty is explicit. Nulls, uncertain, low-confidence all surface clearly.
- Failure modes are visible. You see what the system doesn't know.
Evidence map format
{
"field": "trial_phase",
"value": "Phase III",
"confidence": 0.94,
"evidence": ["...Phase III endpoints met..."],
"span": [142, 168]
}
"field": "trial_phase",
"value": "Phase III",
"confidence": 0.94,
"evidence": ["...Phase III endpoints met..."],
"span": [142, 168]
}
What it is not
Not embeddings-only
Inspectable values, not opaque vectors
Not ontology tooling
Ships consumables, not graph schemas
Not RAG
Governed state, not ephemeral answers
Not “LLM extraction”
Structured pipeline with eval, not prompts