Update feedback processing

app.py

@@ -107,38 +107,45 @@ with gr.Blocks(title="Noteworthy Differences") as demo:
         </table>
 
         """,
-            elem_id="intro_table",
+            elem_id="intro-table",
         )
 
     with gr.Row():
-        with gr.Column():
-            with gr.Accordion("More Info", open=False) as accordion:
+        with gr.Column(scale=1):
+            with gr.Accordion("Query Instructions", open=False) as accordion:
                 gr.Markdown(
-                    """#### Query Instructions
+                    """
                 - Page title is case sensitive; use underscores or spaces
                 - Specify any number of days or up to 499 revisions behind
                   - The closest available revision is retrieved
                 - Only article introductions are downloaded
                 """
                 )
+            with gr.Accordion("Confidence Scores", open=False) as accordion:
                 gr.Markdown(
-                    """#### Confidence Key
+                    """
                     - **High:** heuristic = few-shot, judge agrees
                     - **Moderate:** heuristic ≠ few-shot, judge decides
                     - **Questionable:** heuristic = few-shot, judge vetoes
                     """
                 )
-            random_btn = gr.Button(
-                "Get Random Page Title", size="md", elem_id="random-button"
-            )
-        title_input = gr.Textbox(
-            label="Wikipedia Page Title", placeholder="e.g., Albert Einstein", value=""
-        )
-        number_input = gr.Number(label="Number", value=50, minimum=0, precision=0)
-        units_dropdown = gr.Dropdown(
-            choices=["revisions", "days"], value="revisions", label="Units"
-        )
-        with gr.Column():
+        with gr.Column(scale=3):
+            with gr.Row():
+                page_title = gr.Textbox(
+                    label="Wikipedia Page Title",
+                    placeholder="e.g., Albert Einstein",
+                    value="",
+                )
+                number_behind = gr.Number(
+                    label="Number Behind", value=50, minimum=0, precision=0
+                )
+                units_behind = gr.Dropdown(
+                    choices=["revisions", "days"],
+                    value="revisions",
+                    label="Units Behind",
+                )
+        with gr.Column(scale=1):
+            random_btn = gr.Button("Get Random Page Title", size="md")
             submit_btn = gr.Button(
                 "Fetch Revisions and Run Model", size="md", variant="primary"
             )
@@ -182,8 +189,8 @@ with gr.Blocks(title="Noteworthy Differences") as demo:
                             lines=1,
                             interactive=False,
                         )
-                        confidence = gr.Textbox(
-                            label="Confidence",
+                        confidence_score = gr.Textbox(
+                            label="Confidence Score",
                             lines=1,
                             interactive=False,
                         )
@@ -207,12 +214,12 @@ with gr.Blocks(title="Noteworthy Differences") as demo:
     random_btn.click(
         fn=get_random_wikipedia_title,
         inputs=None,
-        outputs=[title_input],
+        outputs=[page_title],
     )
 
     gr.on(
         # Press Enter in textbox or use button to submit
-        triggers=[title_input.submit, submit_btn.click],
+        triggers=[page_title.submit, submit_btn.click],
         # Clear the new_revision and new_timestamp values before proceeding.
         # The empty values will propagate to the other components (through function return values) if there is an error.
         fn=lambda: (gr.update(value=""), gr.update(value="")),
@@ -222,16 +229,16 @@ with gr.Blocks(title="Noteworthy Differences") as demo:
     ).then(
         # Initialize Logfire context
         fn=start_parent_span,
-        inputs=[title_input, number_input, units_dropdown],
+        inputs=[page_title, number_behind, units_behind],
         outputs=context,
     ).then(
         fn=fetch_current_revision,
-        inputs=[title_input, context],
+        inputs=[page_title, context],
         outputs=[new_revision, new_timestamp],
         api_name=False,
     ).then(
         fn=fetch_previous_revision,
-        inputs=[title_input, number_input, units_dropdown, new_revision, context],
+        inputs=[page_title, number_behind, units_behind, new_revision, context],
         outputs=[old_revision, old_timestamp],
         api_name=False,
     ).then(
@@ -255,7 +262,7 @@ with gr.Blocks(title="Noteworthy Differences") as demo:
             fewshot_rationale,
             context,
         ],
-        outputs=[judge_noteworthy, noteworthy_text, judge_reasoning, confidence],
+        outputs=[judge_noteworthy, noteworthy_text, judge_reasoning, confidence_score],
         api_name=False,
     )
 
@@ -282,7 +289,7 @@ with gr.Blocks(title="Noteworthy Differences") as demo:
             fewshot_rationale,
             context,
         ],
-        outputs=[judge_noteworthy, noteworthy_text, judge_reasoning, confidence],
+        outputs=[judge_noteworthy, noteworthy_text, judge_reasoning, confidence_score],
         api_name=False,
     )
 
@@ -290,9 +297,9 @@ with gr.Blocks(title="Noteworthy Differences") as demo:
     thumbs_up_btn.click(
         fn=save_feedback_agree,
         inputs=[
-            title_input,
-            number_input,
-            units_dropdown,
+            page_title,
+            number_behind,
+            units_behind,
             old_revision,
             new_revision,
             old_timestamp,
@@ -301,7 +308,7 @@ with gr.Blocks(title="Noteworthy Differences") as demo:
             fewshot_rationale,
             judge_reasoning,
             noteworthy_text,
-            confidence,
+            confidence_score,
             heuristic_noteworthy,
             fewshot_noteworthy,
             judge_noteworthy,
@@ -312,9 +319,9 @@ with gr.Blocks(title="Noteworthy Differences") as demo:
     thumbs_down_btn.click(
         fn=save_feedback_disagree,
         inputs=[
-            title_input,
-            number_input,
-            units_dropdown,
+            page_title,
+            number_behind,
+            units_behind,
             old_revision,
             new_revision,
             old_timestamp,
@@ -323,7 +330,7 @@ with gr.Blocks(title="Noteworthy Differences") as demo:
             fewshot_rationale,
             judge_reasoning,
             noteworthy_text,
-            confidence,
+            confidence_score,
             heuristic_noteworthy,
             fewshot_noteworthy,
             judge_noteworthy,
@@ -340,15 +347,11 @@ if __name__ == "__main__":
     head = '<link href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/6.0.0-beta3/css/all.min.css" rel="stylesheet">'
     # Use CSS to style table
     css = """
-    #intro_table {background-color: #eff6ff}
+    #intro-table {background-color: #eff6ff}
     table, tr, td {
         border: none; /* Removes all borders */
         border-collapse: collapse; /* Ensures no gaps between cells */
     }
-
-    #random-button {
-        margin-top: auto; /* Pushes the button to the bottom */
-    }
     """
 
     demo.launch(theme=theme, head=head, css=css)

feedback.py

@@ -6,6 +6,7 @@ import gradio as gr
 import logfire
 import json
 import os
+import re
 
 # Load API keys
 load_dotenv()
@@ -42,9 +43,9 @@ def save_feedback(*args, feedback_value: str) -> None:
     """
     # Assign dict keys to positional arguments
     keys = [
-        "title_input",
-        "number_input",
-        "units_dropdown",
+        "page_title",
+        "number_behind",
+        "units_behind",
         "old_revision",
         "new_revision",
         "old_timestamp",
@@ -53,21 +54,85 @@ def save_feedback(*args, feedback_value: str) -> None:
         "fewshot_rationale",
         "judge_reasoning",
         "noteworthy_text",
-        "confidence",
+        "confidence_score",
         "heuristic_noteworthy",
         "fewshot_noteworthy",
         "judge_noteworthy",
     ]
     feedback_dict = dict(zip(keys, args))
+
+    # Data cleanup
+    # Split dictionary in two
+    keys_start = {"page_title", "number_behind", "units_behind"}
+    dict_start = {k: v for k, v in feedback_dict.items() if k in keys_start}
+    dict_end = {k: v for k, v in feedback_dict.items() if k not in keys_start}
+    # Normalize timestamp and extract revisions behind from display string
+    raw_new_timestamp = dict_end.get("new_timestamp", "")
+    # Extract ISO timestamp (e.g. 2013-12-09T04:48:24Z)
+    ts_match = re.search(r"\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}Z", raw_new_timestamp)
+    if ts_match:
+        dict_end["new_timestamp"] = ts_match.group(0)
+    # Do the same for old timestamp
+    raw_old_timestamp = dict_end.get("old_timestamp", "")
+    ts_match = re.search(r"\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}Z", raw_old_timestamp)
+    if ts_match:
+        dict_end["old_timestamp"] = ts_match.group(0)
+    # Extract revisions behind (e.g. "35 revisions behind")
+    rev_match = re.search(r"(\d+)\s+revisions?\s+behind", raw_old_timestamp)
+    if rev_match:
+        dict_start["revisions_behind"] = int(rev_match.group(1))
+    else:
+        dict_start["revisions_behind"] = None
+    # Merge the start and end dictionaries
+    feedback_dict = dict_start | dict_end
+
     # Add feedback value
     feedback_dict["feedback"] = feedback_value
-    # Save feedback to file
-    feedback_file = f"train-{datetime.now().isoformat()}.json"
-    feedback_path = USER_FEEDBACK_DIR / feedback_file
-    with feedback_path.open("a") as f:
-        f.write(json.dumps(feedback_dict))
-        f.write("\n")
-    gr.Success(f"Saved feedback: <strong>{feedback_value}</strong>", duration=5)
+    do_save = True
+    feedback_action = "Saved"
+
+    # Check for duplicate feedback against the most recent feedback file
+    feedback_files = sorted(
+        USER_FEEDBACK_DIR.glob("*.json"),
+        key=lambda p: p.stat().st_mtime,
+        reverse=True,
+    )
+    if feedback_files:
+        latest_feedback_path = feedback_files[0]
+        with latest_feedback_path.open("r") as latest_f:
+            # Files are written as JSON Lines, but currently contain a single line
+            latest_line = latest_f.readline().strip()
+            if latest_line:
+                latest_feedback = json.loads(latest_line)
+                # Pop the last feedback value
+                old_feedback = latest_feedback.pop("feedback", None)
+                # Use the same feedback value (for detecting resubmission)
+                latest_feedback["feedback"] = feedback_value
+                if latest_feedback == feedback_dict:
+                    # The user resubmitted feedback: remove the previous feedback
+                    latest_feedback_path.unlink()
+                    if feedback_dict["feedback"] == old_feedback:
+                        # The user submitted the same feedback: Issue a warning
+                        gr.Warning(
+                            f"Removed feedback: <strong>{old_feedback}</strong>",
+                            duration=5,
+                        )
+                        do_save = False
+                    else:
+                        # The user changed the feedback: Proceed to update the feedback
+                        feedback_action = "Updated"
+
+    if do_save:
+        # Save feedback to file
+        feedback_file = f"train-{datetime.now().isoformat()}.json"
+        feedback_path = USER_FEEDBACK_DIR / feedback_file
+        with feedback_path.open("a") as f:
+            f.write(json.dumps(feedback_dict))
+            f.write("\n")
+        if feedback_action == "Updated":
+            gr.Info(f"Updated feedback: <strong>{feedback_value}</strong>", duration=5)
+        else:
+            gr.Success(f"Saved feedback: <strong>{feedback_value}</strong>", duration=5)
 
 
 @logfire.instrument("Save feedback: agree")