Create app.py

app.py

@@ -0,0 +1,140 @@
+import os
+import functools
+import re
+from openai import OpenAI
+import tiktoken
+import gradio as gr
+
+# Configure API client
+af_key = os.getenv("OPENAI_API_KEY")
+if not af_key:
+    raise ValueError("Please set the OPENAI_API_KEY environment variable.")
+client = OpenAI(api_key=af_key)
+
+# Available models (env or fallback)
+_env_models = os.getenv("OPENAI_MODEL_LIST", "gpt-3.5-turbo,gpt-4")
+ALL_MODELS = [m.strip() for m in _env_models.split(",") if m.strip()]
+if not ALL_MODELS:
+    ALL_MODELS = ["gpt-3.5-turbo"]
+
+# Token encoder
+@functools.lru_cache(maxsize=64)
+def _get_encoding(model: str):
+    try:
+        return tiktoken.encoding_for_model(model)
+    except KeyError:
+        return tiktoken.get_encoding("cl100k_base")
+
+def count_tokens(text: str, model: str) -> int:
+    enc = _get_encoding(model)
+    return len(enc.encode(text))
+
+# Default banned words (static style list)
+DEFAULT_BANNED_WORDS = [
+    "Hurdles", "Tapestry", "Bustling", "Harnessing", "Unveiling the power",
+    "Realm", "Depicted", "Demistify", "Insurmountable", "New Era",
+    "Poised", "Unravel", "Entanglement", "Unprecedented", "Beacon",
+    "Unleash", "Delve", "Enrich", "Multifaced", "Discover", "Unlock",
+    "Tailored", "Elegant", "Dive", "Ever-evolving", "Adventure",
+    "Journey", "Navigate", "Navigation"
+]
+# Expressions/tones to catch and remove
+DEFAULT_FILTER_PATTERNS = [
+    r"As an AI language model", r"I(?:’|'|i)m sorry", r"I apologize",
+    r"In conclusion", r"At the end of the day"
+]
+
+# Original style prefix
+ORIGINAL_PREFIX = (
+    "You are a clear-writing assistant. Follow these rules for every response:\n"
+    "1. Write in plain, everyday language.\n"
+    "2. Avoid fluff, jargon, and tautological expressions.\n"
+    "3. Use short, direct sentences.\n"
+    "4. Do not use complicated English words—choose simple alternatives.\n"
+    "5. Do not use any of the default banned words.\n"
+    "Whenever you’d normally reach for a banned word, pick a simple synonym or rephrase.\n"
+)
+
+# Additional prompting points
+ADDITIONAL_POINTS = {
+    "Tone & Voice": [
+        "Adopt a confident but approachable tone.",
+        "Write as if you’re explaining to a colleague.",
+        "Use active voice and vary sentence length."
+    ],
+    "Audience & Purpose": [
+        "Assume the reader has a graduate-level background.",
+        "Focus on practical takeaways for a project manager.",
+        "Start with an executive summary, then technical details."
+    ],
+    "Structure & Signposting": [
+        "Begin with a 2-sentence overview, then clear headings.",
+        "Provide an outline before details.",
+        "End with three bullet-point recommendations."
+    ]
+}
+
+# Build the Gradio interface
+with gr.Blocks() as demo:
+    gr.Markdown("# Custom Prompt Builder for LLMs")
+
+    model_dropdown = gr.Dropdown(ALL_MODELS, value=ALL_MODELS[0], label="Select Model")
+    user_query = gr.Textbox(label="User Query", placeholder="Type your question here...", lines=3)
+
+    with gr.Accordion("Original Style Instructions", open=False):
+        gr.Textbox(value=ORIGINAL_PREFIX, interactive=False, lines=8)
+
+    # Checkboxes for additional points
+    point_widgets = []
+    for category, pts in ADDITIONAL_POINTS.items():
+        widget = gr.CheckboxGroup(pts, label=category)
+        point_widgets.append(widget)
+
+    # Input for extra banned expressions (regex patterns)
+    custom_bans = gr.Textbox(
+        label="Custom banned expressions (comma-separated regex)", lines=2,
+        placeholder="e.g. 'As an AI language model','I apologize'"
+    )
+
+    # Generate and display the AI response
+    def build_and_respond(query, model, *args):
+        # Last arg is custom bans, rest are selected points
+        *selected_lists, custom_raw = args
+        selections = [item for sublist in selected_lists for item in sublist]
+
+        # Construct full prefix
+        style_block = ORIGINAL_PREFIX + "\n" + "\n".join(selections)
+        full_prompt = style_block + "\n\nNow, answer the user’s question.\n\n" + query
+
+        # Call the OpenAI API
+        resp = client.chat.completions.create(
+            model=model,
+            messages=[{"role": "user", "content": full_prompt}]
+        )
+        reply = resp.choices[0].message.content
+
+        # Build complete banned list
+        custom_list = [b.strip() for b in custom_raw.split(",") if b.strip()]
+        banned_words = DEFAULT_BANNED_WORDS
+        banned_patterns = DEFAULT_FILTER_PATTERNS + custom_list
+
+        # Remove banned words and patterns
+        for word in banned_words:
+            reply = re.sub(rf"\b{re.escape(word)}\b", "", reply, flags=re.IGNORECASE)
+        for pat in banned_patterns:
+            reply = re.sub(pat, "", reply, flags=re.IGNORECASE)
+
+        # Clean extra whitespace
+        reply = re.sub(r"\s{2,}", " ", reply).strip()
+
+        # Token usage
+        tokens_used = count_tokens(full_prompt, model) + count_tokens(reply, model)
+        return reply + f"\n\n(Tokens used: {tokens_used})"
+
+    inputs = [user_query, model_dropdown] + point_widgets + [custom_bans]
+    output = gr.Chatbot(label="AI Response")
+
+    user_query.submit(build_and_respond, inputs=inputs, outputs=output)
+    user_query.submit(lambda: "", None, user_query)
+
+    demo.launch()