Review page size option, fixing mkdirs in convertsilver script

pdelfin/eval/runeval.py

@@ -180,7 +180,7 @@ def process_jsonl_file(jsonl_file, gold_data, comparer):
 
     return total_alignment_score, char_weighted_alignment_score, total_chars, total_pages, page_data
 
-def do_eval(gold_data_path: str, eval_data_path: str, review_page_name: str) -> tuple[float, list[dict]]:
+def do_eval(gold_data_path: str, eval_data_path: str, review_page_name: str, review_page_size: int) -> tuple[float, list[dict]]:
     gold_data = load_gold_data(gold_data_path)
 
     total_alignment_score = 0
@@ -237,10 +237,10 @@ def do_eval(gold_data_path: str, eval_data_path: str, review_page_name: str) ->
 
     # Select the top 20 lowest alignments
     page_eval_data.sort(key=lambda x: x["alignment"])
-    create_review_html(page_eval_data[:20], filename=review_page_name + "_worst.html")
+    create_review_html(page_eval_data[:review_page_size], filename=review_page_name + "_worst.html")
 
     # Select random entries to return in the page_eval_data
-    page_eval_data = random.sample(page_eval_data, 20)
+    page_eval_data = random.sample(page_eval_data, review_page_size)
     create_review_html(page_eval_data, filename=review_page_name + "_sample.html")
 
 
@@ -256,6 +256,12 @@ def do_eval(gold_data_path: str, eval_data_path: str, review_page_name: str) ->
         default="review_page",
         help="What name to give to this evaluation/comparison"
     )
+    parser.add_argument(
+        '--review_size',
+        default=20,
+        type=int,
+        help="Number of entries to show on the generated review page",
+    )
     parser.add_argument(
         'gold_data_path',
         type=str,
@@ -269,4 +275,4 @@ def do_eval(gold_data_path: str, eval_data_path: str, review_page_name: str) ->
 
     args = parser.parse_args()
 
-    result = do_eval(gold_data_path=args.gold_data_path, eval_data_path=args.eval_data_path, review_page_name=args.name)
+    result = do_eval(gold_data_path=args.gold_data_path, eval_data_path=args.eval_data_path, review_page_name=args.name, review_page_size=args.review_size)

pdelfin/prompts/anchor.py

@@ -36,17 +36,21 @@ def get_anchor_text(local_pdf_path: str, page: int, pdf_engine: Literal["pdftote
     elif pdf_engine == "pymupdf":
         return _get_pymupdf(local_pdf_path, page)
     elif pdf_engine == "topcoherency":
-        options = [
-            _get_pdftotext(local_pdf_path, page),
-            _get_pymupdf(local_pdf_path, page),
-            _get_pdfium(local_pdf_path, page),
-            _get_pypdf_raw(local_pdf_path, page)
-        ]
+        options = {
+            "pdftotext": _get_pdftotext(local_pdf_path, page),
+            "pymupdf": _get_pymupdf(local_pdf_path, page),
+            "pdfium": _get_pdfium(local_pdf_path, page),
+            "pypdf_raw": _get_pypdf_raw(local_pdf_path, page)
+        }
 
-        scores = [get_document_coherency(text) for text in options]
+        scores = {label: get_document_coherency(text) for label, text in options.items()}
 
-        # return option with the best (highest) score (higher is more likley, as these are logprobs)
-        return options[scores.index(max(scores))]
+        best_option_label = max(scores, key=scores.get)
+        best_option = options[best_option_label]
+
+        print(f"topcoherency chosen: {best_option_label}")
+
+        return best_option
     elif pdf_engine == "pdfreport":
         return _linearize_pdf_report(_pdf_report(local_pdf_path, page))
     else:

pdelfin/silver_data/convertsilver_openai.py

@@ -4,6 +4,7 @@
 from pathlib import Path
 from concurrent.futures import ProcessPoolExecutor, as_completed
 import sys
+import os
 import logging
 
 import smart_open
@@ -63,6 +64,25 @@ def process_file(input_file: str, output_file: str, rewrite_prompt_str: bool):
 
                     if match:
                         raw_page_text = match.group(1).strip()
+
+                        # Ok, now we want to try to see if it's better if we recalculate the anchor text
+                        goldkey = obj["custom_id"]
+                        s3_path = goldkey[:goldkey.rindex("-")]
+                        page = int(goldkey[goldkey.rindex("-") + 1:])
+
+                        # Save the pdf to a temporary cache folder
+                        import os
+                        local_pdf_path = "/home/ubuntu/.cache/samplepdfs/" + os.path.basename(s3_path)
+
+                        if not os.path.exists(local_pdf_path):
+                            print("Loading pdf", s3_path)
+                            with smart_open.smart_open(s3_path, "rb") as fin, open(local_pdf_path, "wb") as fout:
+                                fout.write(fin.read())
+
+                        from pdelfin.prompts.anchor import get_anchor_text
+
+                        raw_page_text = get_anchor_text(local_pdf_path, page, pdf_engine="topcoherency")
+
                         from pdelfin.prompts import build_openai_silver_data_prompt
                         obj["body"]["messages"][0]["content"][0]["text"] = build_openai_silver_data_prompt(raw_page_text)
 
@@ -74,6 +94,7 @@ def process_file(input_file: str, output_file: str, rewrite_prompt_str: bool):
 
         logging.info(f"Processed '{input_file}': {processed_count} records transformed, {error_count} errors.")
     except Exception as e:
+        logging.exception(e)
         logging.error(f"Failed to process file {input_file}: {e}")
 
 
@@ -191,6 +212,9 @@ def main():
     output_dir = args.output_dir.rstrip('/')
     max_jobs = args.jobs
 
+    if not output_dir.startswith("s3:"):
+        os.makedirs(output_dir, exist_ok=True)
+
     # List input files
     input_files = list_input_files(input_dir)