Default top_n to None, strip metadata row, post-filter by min_conf

main.py

@@ -22,6 +22,13 @@ def health():
     }
 
 
+# Keys we expect on a real detection row — used to filter out the
+# metadata block BirdNET appends to its CSV output ("File splitting
+# duration", "Segment length", etc., which DictReader otherwise picks
+# up as a phantom row).
+DETECTION_KEYS = {"Start (s)", "End (s)", "Scientific name", "Common name", "Confidence"}
+
+
 @app.post("/analyze")
 async def analyze_audio(
     audio: UploadFile = File(...),
@@ -29,7 +36,10 @@ async def analyze_audio(
     lat: Optional[float] = Form(None),
     lon: Optional[float] = Form(None),
     week: Optional[int] = Form(None),
-    top_n: int = Form(5),
+    # None lets min_conf gate the results. Pass an int to force "top N per
+    # segment regardless of confidence" — useful for debugging silent files,
+    # noisy for normal use.
+    top_n: Optional[int] = Form(None),
 ):
     suffix = Path(audio.filename or "audio.wav").suffix or ".wav"
     with tempfile.NamedTemporaryFile(suffix=suffix, delete=False) as tmp_in:
@@ -54,6 +64,17 @@ async def analyze_audio(
         for csv_file in Path(out_dir).glob("*.csv"):
             with open(csv_file, newline="") as f:
                 for row in csv_module.DictReader(f):
+                    # Skip the trailing metadata block — its keys don't
+                    # overlap with detection keys.
+                    if not DETECTION_KEYS.issubset(row.keys()):
+                        continue
+                    # Defence-in-depth: drop rows below threshold even
+                    # when top_n is set, so the response is consistent.
+                    try:
+                        if float(row.get("Confidence", 0)) < min_conf:
+                            continue
+                    except (TypeError, ValueError):
+                        continue
                     results.append(dict(row))
 
         return JSONResponse({"results": results})