Input text — raw and unprocessed
Instruction and examples — your examples define the schema
Chunk the input — long text splits into overlapping windows
Extract in parallel — every chunk carries the same prompt
Extractions — merged and aligned
Source grounding — every value maps to its exact span
input text
Patient started on 40mg lisinopril
daily. BP 128/82 today.
Active problem: hypertension
Interval symptom: mild dizziness
since Tuesday. Recheck in 4 weeks.
prompt
chunks + prompt
model
extractions · merged & aligned
instruction
Extract medications, conditions
and symptoms from the text.
few-shot examples — these define the schema
"aspirin 81mg" → medication · dosage
"low-grade fever" → symptom · onset
…
chunk 1 chars 0–64
chunk 2 chars 48–112
chunk 3 chars 96–155
chunk 1
chunk 2
chunk 3
prompt
LLM
parallel passes
medication
lisinopril
dosage: 40mg · freq: daily · chars 24–34
condition
hypertension
status: active · chars 75–87
symptom
mild dizziness
onset: Tuesday · chars 106–120