Change ocr prompt to instead enhance tesseract hOCR

Dockerfile

@@ -25,7 +25,7 @@ RUN adduser -S -G nobody scyllaridae
 
 RUN apk update && \
     apk add --no-cache \
-      curl==8.11.0-r2 \
+      curl==8.11.1-r0 \
       bash==5.2.26-r0 \
       ca-certificates==20240705-r0 \
       openssl==3.3.2-r1

examples/openai-htr/Dockerfile

@@ -12,9 +12,8 @@ RUN --mount=type=bind,from=imagemagick,source=/packages,target=/packages \
 
 RUN magick -list format | grep "JPEG-2000"
 
-ENV OPENAI_MODEL=gpt-4o-mini \
-    PROMPT="Transcribe this image that contains handwritten text. Include all text you see in the image. In your response, say absolutely nothing except the text from the image." \
-    MAX_TOKENS=300
+ENV OPENAI_MODEL=gpt-4o \
+    CHAT_PROMPT="changeme" \
+    MAX_TOKENS=16384
 
-COPY scyllaridae.yml /app/scyllaridae.yml
-COPY cmd.sh /app/cmd.sh
+COPY . /app/

examples/openai-htr/cmd.sh

@@ -3,10 +3,26 @@
 set -eou pipefail
 
 TMP_DIR=$(mktemp -d)
+HOCR_URL="$1"
+DOMAIN=$(echo "$HOCR_URL"| awk -F/ '{print $1"//"$3}')
+# our hOCR filenames are the node ID
+NID=$(echo "$HOCR_URL" | xargs basename | awk -F '.hocr' '{print $1}')
 
+# take the base prompt and move it into place
+cp /app/prompt.txt "$TMP_DIR/prompt.txt"
+
+# the hOCR document is being streamed into this script
+# append the hOCR document into the prompt
+# since we're asking the LLM to improve the hOCR doc
+cat >> "$TMP_DIR/prompt.txt"
+CHAT_PROMPT=$(jq --null-input --rawfile rawstring "$TMP_DIR/prompt.txt" '$rawstring')
+
+# find the service file
+SERVICE_FILE_PATH=$(curl -s "$DOMAIN/node/$NID/service-file" | jq -r '.[0].file')
 # convert service file to jpg
-magick - "$TMP_DIR/img.jpg"
+curl -s "${DOMAIN}${SERVICE_FILE_PATH}" | magick - "$TMP_DIR/img.jpg"
 
+# chatgpt needs it base64 encoded
 BASE64_IMAGE=$(base64 -w 0 "$TMP_DIR/img.jpg")
 
 cat <<EOF > "$TMP_DIR/payload.json"
@@ -18,7 +34,7 @@ cat <<EOF > "$TMP_DIR/payload.json"
       "content": [
         {
           "type": "text",
-          "text": "$PROMPT"
+          "text": $CHAT_PROMPT
         },
         {
           "type": "image_url",

examples/openai-htr/prompt.txt

@@ -0,0 +1,8 @@
+Transcribe the text in the provided image.
+Place the transcribed text into the appropriate ocrx_word span in the hOCR template provided. Ensure the following:
+Never use tesseract to aid you.
+Include all text in the transcription.
+Never return an incomplete document. Always finish the complete job, with every word wrapped in a span, and the words accurately transcribed from the image.
+Never edit the hOCR dimensions provided. Only replace the text within <span>'s. 
+Do not wrap your response in backticks. Your response should only be the hOCR document provided below with your transcribed text placed in the appropriate spots in the hOCR document.
+Do not add any additional spans or attributes. You only should be adding plain text within the hocr HTML document.

examples/openai-htr/scyllaridae.yml

@@ -1,5 +1,9 @@
 allowedMimeTypes:
-  - "image/*"
+  - "text/vnd.hocr+html"
+  - "text/html"
+  - "application/octet-stream"
 cmdByMimeType:
   default:
     cmd: /app/cmd.sh
+    args:
+      - "%source-uri"

internal/config/server.go

@@ -105,6 +105,9 @@ func IsAllowedMimeType(mimetype string, allowedFormats []string) bool {
 				return true
 			}
 		}
+		if format == strings.Split(mimetype, ";")[0] {
+			return true
+		}
 	}
 	return false
 }