Long and large documents
Does it hold up on long documents and long field lists?
Length: the same documents at about 1,000, 4,000, 16,000 and 48,000 tokens, padded with similar documents, up to each model's limit.
Fields: the same 5 fields asked alone, inside 20 and inside 50.
Long-document drop points lost from 1k to 48k tokens
Clairic
Cloudflare
NuExtract3
NuMind
GLiNER2
Fastino
Accuracy at about 1,000 tokens minus accuracy at about 48,000 tokens, on the same documents padded with similar ones.
Many-fields drop points lost from 5 to 50 fields
Clairic
Cloudflare
NuExtract3
NuMind
GLiNER2
Fastino
Accuracy on the same 5 fields when asked alone versus inside a list of 50.
Test items and precision
- Items
- 60 documents at four lengths; 60 documents with 5, 20 and 50 fields
- Precision
- About ±6 points per step; the trend across steps is the main result.