← Back to V2 Schema Changes
V2 Schema Changes Record
Page Classifications · V2 Schema Changes · #13
id
13
ts
1778851055
kind
observation
before
BGE text-embedding NN for handwritten
after
~50% precision on high-signal (2+ embeddings)
rationale
Text embeddings can't reliably distinguish handwritten from photo/blank because OCR text is uniformly noisy for all three. For better handwritten discovery, use: (1) CLIP visual embeddings (611K already indexed via /find-image), (2) Bates-range search within known investigation series (MCC notes form a contiguous range 135084-135584), (3) direct vision classification of v1=image.scan candidates.
n docs touched
0