WIP

scripts/analyze_v3.py

@@ -0,0 +1,103 @@
+#!/usr/bin/env python3
+"""Analyze batch inference v3 results (Round 2 fixes)."""
+
+import json
+from collections import Counter
+
+with open("scripts/inference_results_v3.json") as f:
+    results = json.load(f)
+
+total = len(results)
+success = sum(1 for r in results if r["status"] == 200)
+print(f"Total PDFs: {total}, Successful: {success}")
+print()
+
+# Summary table
+header = f"{'PDF':<40} {'Det':<4} {'Fld':<4} {'Time':<7} Fields"
+print(header)
+print("-" * 140)
+for r in results:
+    fn = r["filename"][:39]
+    data = r.get("data", {})
+    result_data = data.get("result", {})
+    fields = result_data.get("fields", {})
+    dets = len(result_data.get("detections", []))
+    nfields = len(fields)
+    t = r["time_seconds"]
+    parts = []
+    for k, v in fields.items():
+        sv = str(v)
+        if len(sv) > 30:
+            sv = sv[:27] + "..."
+        parts.append(f"{k}={sv}")
+    field_str = ", ".join(parts)
+    print(f"{fn:<40} {dets:<4} {nfields:<4} {t:<7} {field_str}")
+
+print()
+
+# Field coverage
+field_counts: Counter = Counter()
+conf_sums: Counter = Counter()
+ok_count = 0
+for r in results:
+    if r["status"] != 200:
+        continue
+    ok_count += 1
+    result_data = r["data"]["result"]
+    for k in result_data.get("fields", {}):
+        field_counts[k] += 1
+    for k, v in (result_data.get("confidence") or {}).items():
+        conf_sums[k] += v
+
+print(f"Field Coverage ({ok_count} successful PDFs):")
+hdr = f"{'Field':<35} {'Present':<10} {'Rate':<10} {'Avg Conf':<10}"
+print(hdr)
+print("-" * 65)
+for field in [
+    "InvoiceNumber", "InvoiceDate", "InvoiceDueDate", "OCR",
+    "Amount", "Bankgiro", "Plusgiro",
+    "supplier_organisation_number", "customer_number", "payment_line",
+]:
+    cnt = field_counts.get(field, 0)
+    rate = cnt / ok_count * 100 if ok_count else 0
+    avg_conf = conf_sums.get(field, 0) / cnt if cnt else 0
+    flag = ""
+    if rate < 30:
+        flag = " <<<"
+    elif rate < 60:
+        flag = " !!"
+    print(f"{field:<35} {cnt:<10} {rate:<10.1f} {avg_conf:<10.3f}{flag}")
+
+# Fallback count
+fb_count = 0
+for r in results:
+    if r["status"] == 200:
+        result_data = r["data"]["result"]
+        if result_data.get("fallback_used"):
+            fb_count += 1
+print(f"\nFallback used: {fb_count}/{ok_count}")
+
+# Low-confidence fields
+print("\nLow-confidence extractions (< 0.7):")
+for r in results:
+    if r["status"] != 200:
+        continue
+    result_data = r["data"]["result"]
+    for k, v in (result_data.get("confidence") or {}).items():
+        if v < 0.7:
+            fv = result_data.get("fields", {}).get(k, "?")
+            print(f"  [{v:.3f}] {k:<25} = {str(fv)[:40]:<40} ({r['filename'][:36]})")
+
+# PDFs with very few fields (possible issues)
+print("\nPDFs with <= 2 fields extracted:")
+for r in results:
+    if r["status"] != 200:
+        continue
+    result_data = r["data"]["result"]
+    fields = result_data.get("fields", {})
+    if len(fields) <= 2:
+        print(f"  {r['filename']}: {len(fields)} fields - {list(fields.keys())}")
+
+# Avg time
+avg_time = sum(r["time_seconds"] for r in results) / len(results)
+print(f"\nAverage processing time: {avg_time:.2f}s")