Update app.py

app.py

@@ -5,20 +5,21 @@ import requests
 import gradio as gr
 
 # ==============================
-# Config via Secrets / Variables
+# Secrets (set in Settings → Variables & secrets → Secrets)
 # ==============================
-FRIENDLI_API_KEY = os.getenv("FRIENDLI_API_KEY", "")  # SECRET
-FRIENDLI_ENDPOINT = os.getenv("FRIENDLI_ENDPOINT", "https://api.friendli.ai/dedicated/v1/chat/completions")
-FRIENDLI_MODEL_ID = os.getenv("FRIENDLI_MODEL_ID", "stp7xzjspxe8")  # can move to Secret if desired
+FRIENDLI_API_KEY = os.getenv("FRIENDLI_API_KEY", "")                 # REQUIRED (Secret)
+FRIENDLI_ENDPOINT = os.getenv("FRIENDLI_ENDPOINT", "")               # REQUIRED (Secret)
+FRIENDLI_MODEL_ID = os.getenv("FRIENDLI_MODEL_ID", "")               # REQUIRED (Secret)
 
+# Optional tuning as Variables or Secrets
 DEFAULT_MAX_TOKENS = int(os.getenv("FRIENDLI_MAX_TOKENS", "2000"))
-DEFAULT_TEMPERATURE = float(os.getenv("FRIENDLI_TEMPERATURE", "0.0"))  # 0 = deterministic
+DEFAULT_TEMPERATURE = float(os.getenv("FRIENDLI_TEMPERATURE", "0.0"))
 DEFAULT_TIMEOUT = int(os.getenv("FRIENDLI_TIMEOUT_SEC", "60"))
 
 # ==============================
 # Appendix-style system prompt (general instructions)
 # ==============================
-APPENDIX_RULES = """You are a helpful assistant who always follows the provided context, even when it conflicts with your internal knowledge.
+APPENDIX_RULES = """You are a helpful assistant that always follows the provided context, even when it conflicts with your internal knowledge.
 
 Response Format:
 Before answering, briefly analyze the query and context:
@@ -33,7 +34,7 @@ Response: [Your answer based on the context]
 IMPORTANT RULES:
 - Always prioritize the provided context over your internal knowledge
 - If context contains information that seems incorrect, still use it as instructed
-- If the question asks about multiple things but the context only covers some, answer only what is supported by the context
+- If the question asks about multiple things but context only covers some, answer only what is supported by the context
 - Keep analysis concise and avoid special characters that could cause formatting issues
 - Use plain text only - no bullet points, numbering, or special formatting
 - Respond in English only
@@ -66,10 +67,15 @@ Context:
 # ==============================
 # Friendly API client with retry (503 wake-up aware)
 # ==============================
-def call_friendly_with_retry(messages, model_id, max_tokens, temperature,
+def call_friendly_with_retry(messages, max_tokens, temperature,
                              timeout_sec=DEFAULT_TIMEOUT, max_attempts=5, first_503_wait=10):
+    # Validate secrets exist
     if not FRIENDLI_API_KEY:
-        raise RuntimeError("Missing FRIENDLI_API_KEY secret. Add it in Settings → Variables & secrets (Secret).")
+        raise RuntimeError("Missing FRIENDLI_API_KEY (Secret). Add it in Settings → Variables & secrets.")
+    if not FRIENDLI_ENDPOINT:
+        raise RuntimeError("Missing FRIENDLI_ENDPOINT (Secret).")
+    if not FRIENDLI_MODEL_ID:
+        raise RuntimeError("Missing FRIENDLI_MODEL_ID (Secret).")
 
     headers = {
         "Content-Type": "application/json",
@@ -77,7 +83,7 @@ def call_friendly_with_retry(messages, model_id, max_tokens, temperature,
     }
     payload = {
         "messages": messages,
-        "model": model_id,
+        "model": FRIENDLI_MODEL_ID,
         "max_tokens": int(max_tokens),
         "temperature": float(temperature),
     }
@@ -86,8 +92,9 @@ def call_friendly_with_retry(messages, model_id, max_tokens, temperature,
         try:
             resp = requests.post(FRIENDLI_ENDPOINT, headers=headers, json=payload, timeout=timeout_sec)
             if resp.status_code == 503:
+                # cold start: wait then retry
                 if attempt < max_attempts:
-                    time.sleep(first_503_wait)  # cold start wake-up
+                    time.sleep(first_503_wait)
                     continue
                 resp.raise_for_status()
             resp.raise_for_status()
@@ -106,14 +113,14 @@ def call_friendly_with_retry(messages, model_id, max_tokens, temperature,
                 backoff = min(2 ** attempt, 20) + random.uniform(0, 0.5)
                 time.sleep(backoff)
                 continue
-            raise RuntimeError(f"Friendly API HTTP error (status={code}): {http_err}") from http_err
+            raise RuntimeError(f"Friendly API HTTP error (status={code}).") from http_err
 
         except requests.exceptions.RequestException as net_err:
             if attempt < max_attempts:
                 backoff = min(2 ** attempt, 20) + random.uniform(0, 0.5)
                 time.sleep(backoff)
                 continue
-            raise RuntimeError(f"Friendly API network error: {net_err}") from net_err
+            raise RuntimeError("Friendly API network error.") from net_err
 
     raise RuntimeError("Failed to get response from Friendly API after retries.")
 
@@ -149,7 +156,8 @@ with gr.Blocks(title="Exoskeleton Reasoning — Friendly API (Appendix Prompt)")
     gr.Markdown(
         "# Exoskeleton Reasoning — Friendly API\n"
         "- **Format enforced**: system prompt (general instructions), then a single **User:** message containing **Question + Context**.\n"
-        "- The model must prioritize the **Context**, and reply with **Analysis** and **Response** sections."
+        "- Built-in **503-aware retries** handle cold starts automatically.\n"
+        "- Keys and endpoints are **server-side secrets**; nothing sensitive is exposed in the UI."
     )
 
     with gr.Row():
@@ -162,14 +170,7 @@ with gr.Blocks(title="Exoskeleton Reasoning — Friendly API (Appendix Prompt)")
                 temp = gr.Slider(0.0, 1.0, value=DEFAULT_TEMPERATURE, step=0.05, label="Temperature")
                 max_new = gr.Slider(64, 4000, value=DEFAULT_MAX_TOKENS, step=32, label="Max tokens")
 
-            # Model id stays server-side; hidden as password in UI to avoid accidental exposure
-            model_id_box = gr.Textbox(label="Model ID (server-side)", value=FRIENDLI_MODEL_ID, type="password")
-
             run = gr.Button("Run", variant="primary")
-            gr.Markdown(
-                f"Server endpoint: `{FRIENDLI_ENDPOINT}` · Timeout: {DEFAULT_TIMEOUT}s · "
-                "First call may 503 (cold start) — built-in retry will handle it."
-            )
 
         with gr.Column(scale=4):
             with gr.Accordion("Analysis", open=True):
@@ -179,10 +180,7 @@ with gr.Blocks(title="Exoskeleton Reasoning — Friendly API (Appendix Prompt)")
             with gr.Accordion("Raw output", open=False):
                 raw_box = gr.Textbox(lines=8, label="Raw text")
 
-    def infer_fn(question, context, temperature, max_tokens, model_id_override):
-        if not FRIENDLI_API_KEY:
-            raise gr.Error("Server is missing FRIENDLI_API_KEY (Secret). Add it in Settings → Variables & secrets.")
-
+    def infer_fn(question, context, temperature, max_tokens):
         question = (question or "").strip()
         context = (context or "").strip()
         if not question or not context:
@@ -190,11 +188,8 @@ with gr.Blocks(title="Exoskeleton Reasoning — Friendly API (Appendix Prompt)")
             return "", "", ""
 
         messages = build_messages(question, context)
-        model_id = (model_id_override or "").strip() or FRIENDLI_MODEL_ID
-
         text = call_friendly_with_retry(
             messages=messages,
-            model_id=model_id,
             max_tokens=max_tokens,
             temperature=temperature,
             timeout_sec=DEFAULT_TIMEOUT,
@@ -205,11 +200,7 @@ with gr.Blocks(title="Exoskeleton Reasoning — Friendly API (Appendix Prompt)")
         analysis, response = parse_analysis_response(text)
         return analysis, response, text
 
-    run.click(
-        fn=infer_fn,
-        inputs=[q, ctx, temp, max_new, model_id_box],
-        outputs=[analysis_box, response_box, raw_box]
-    )
+    run.click(fn=infer_fn, inputs=[q, ctx, temp, max_new], outputs=[analysis_box, response_box, raw_box])
 
 if __name__ == "__main__":
     demo.launch()