Update app.py
Browse files
app.py
CHANGED
|
@@ -33,7 +33,12 @@ GROQ_API_KEY = os.getenv("GROQ_API_KEY")
|
|
| 33 |
GROQ_MODEL = os.getenv("GROQ_MODEL") # optional override
|
| 34 |
OPENROUTER_API_KEY = os.getenv("OPENROUTER_API_KEY")
|
| 35 |
OPENROUTER_MODEL = os.getenv("OPENROUTER_MODEL", "meta-llama/llama-3.3-70b-instruct:free")
|
| 36 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
|
| 38 |
|
| 39 |
# ----------------------------------------------------------------- TOOLS
|
|
@@ -160,6 +165,39 @@ def youtube_transcript(url: str) -> str:
|
|
| 160 |
|
| 161 |
|
| 162 |
# ----------------------------------------------------------------- AGENT
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 163 |
def pick_groq_model() -> str:
|
| 164 |
"""Ask Groq which models this key can use and pick the strongest chat model."""
|
| 165 |
if GROQ_MODEL:
|
|
@@ -184,7 +222,22 @@ def pick_groq_model() -> str:
|
|
| 184 |
|
| 185 |
|
| 186 |
def build_model():
|
| 187 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 188 |
if GROQ_API_KEY:
|
| 189 |
return LiteLLMModel(
|
| 190 |
model_id=f"groq/{pick_groq_model()}",
|
|
@@ -201,12 +254,11 @@ def build_model():
|
|
| 201 |
)
|
| 202 |
if GEMINI_API_KEY:
|
| 203 |
return LiteLLMModel(
|
| 204 |
-
model_id=f"gemini/{
|
| 205 |
api_key=GEMINI_API_KEY,
|
| 206 |
-
timeout=
|
| 207 |
-
num_retries=
|
| 208 |
)
|
| 209 |
-
'''
|
| 210 |
return InferenceClientModel(
|
| 211 |
model_id=os.getenv("MODEL_ID", "Qwen/Qwen2.5-72B-Instruct"), token=HF_TOKEN
|
| 212 |
)
|
|
@@ -227,7 +279,7 @@ class BasicAgent:
|
|
| 227 |
self.agent = CodeAgent(
|
| 228 |
tools=[
|
| 229 |
DuckDuckGoSearchTool(),
|
| 230 |
-
VisitWebpageTool(),
|
| 231 |
download_task_file,
|
| 232 |
read_file,
|
| 233 |
transcribe_audio,
|
|
@@ -300,7 +352,7 @@ def run_and_submit_all(profile: gr.OAuthProfile | None = None):
|
|
| 300 |
print(f"Task {task_id} failed: {err}")
|
| 301 |
answers.append({"task_id": task_id, "submitted_answer": ans})
|
| 302 |
log.append({"Task ID": task_id, "Question": question[:80], "Submitted Answer": ans, "Error": err})
|
| 303 |
-
time.sleep(
|
| 304 |
|
| 305 |
if not any(a["submitted_answer"] for a in answers):
|
| 306 |
first_err = next((l["Error"] for l in log if l["Error"]), "unknown")
|
|
@@ -334,7 +386,7 @@ def test_one():
|
|
| 334 |
try:
|
| 335 |
agent = BasicAgent(max_steps=3)
|
| 336 |
q = ".rewsna eht sa \"tfel\" drow eht fo etisoppo eht etirw ,ecnetnes siht dnatsrednu uoy fI"
|
| 337 |
-
used =
|
| 338 |
return f"Model: {used}\nAnswer: {agent(q, 'test', '')} (expected: right)"
|
| 339 |
except Exception as e:
|
| 340 |
return f"FAILED: {type(e).__name__}: {e}"
|
|
|
|
| 33 |
GROQ_MODEL = os.getenv("GROQ_MODEL") # optional override
|
| 34 |
OPENROUTER_API_KEY = os.getenv("OPENROUTER_API_KEY")
|
| 35 |
OPENROUTER_MODEL = os.getenv("OPENROUTER_MODEL", "meta-llama/llama-3.3-70b-instruct:free")
|
| 36 |
+
CEREBRAS_API_KEY = os.getenv("CEREBRAS_API_KEY")
|
| 37 |
+
MISTRAL_API_KEY = os.getenv("MISTRAL_API_KEY")
|
| 38 |
+
ANY_KEY = bool(
|
| 39 |
+
CEREBRAS_API_KEY or MISTRAL_API_KEY or GROQ_API_KEY
|
| 40 |
+
or OPENROUTER_API_KEY or GEMINI_API_KEY or HF_TOKEN
|
| 41 |
+
)
|
| 42 |
|
| 43 |
|
| 44 |
# ----------------------------------------------------------------- TOOLS
|
|
|
|
| 165 |
|
| 166 |
|
| 167 |
# ----------------------------------------------------------------- AGENT
|
| 168 |
+
def pick_compat_model(url: str, key: str, prefs: tuple, extra_skip: tuple = ()) -> str:
|
| 169 |
+
"""List models from an OpenAI-compatible provider and pick the best chat model."""
|
| 170 |
+
r = requests.get(url, headers={"Authorization": f"Bearer {key}"}, timeout=30)
|
| 171 |
+
r.raise_for_status()
|
| 172 |
+
ids = [m["id"] for m in r.json().get("data", [])]
|
| 173 |
+
skip = ("whisper", "guard", "tts", "embed", "moderation", "ocr", "voxtral", "transcribe") + tuple(extra_skip)
|
| 174 |
+
chat = [i for i in ids if not any(s in i.lower() for s in skip)]
|
| 175 |
+
for p in prefs:
|
| 176 |
+
for i in chat:
|
| 177 |
+
if p in i.lower():
|
| 178 |
+
return i
|
| 179 |
+
if chat:
|
| 180 |
+
return chat[0]
|
| 181 |
+
raise RuntimeError(f"No usable chat model found. Models seen: {ids}")
|
| 182 |
+
|
| 183 |
+
|
| 184 |
+
def gemini_model_name() -> str:
|
| 185 |
+
"""Use GEMINI_MODEL if set, else ask Google which Flash model this key can use."""
|
| 186 |
+
if os.getenv("GEMINI_MODEL"):
|
| 187 |
+
return os.getenv("GEMINI_MODEL")
|
| 188 |
+
try:
|
| 189 |
+
m = pick_compat_model(
|
| 190 |
+
"https://generativelanguage.googleapis.com/v1beta/openai/models",
|
| 191 |
+
GEMINI_API_KEY,
|
| 192 |
+
("3.8-flash", "3.5-flash", "flash"),
|
| 193 |
+
extra_skip=("lite", "image", "live", "audio", "robotics", "computer", "imagen", "veo"),
|
| 194 |
+
)
|
| 195 |
+
return m.replace("models/", "")
|
| 196 |
+
except Exception as e:
|
| 197 |
+
print(f"Gemini model auto-pick failed, using default: {e}")
|
| 198 |
+
return GEMINI_MODEL
|
| 199 |
+
|
| 200 |
+
|
| 201 |
def pick_groq_model() -> str:
|
| 202 |
"""Ask Groq which models this key can use and pick the strongest chat model."""
|
| 203 |
if GROQ_MODEL:
|
|
|
|
| 222 |
|
| 223 |
|
| 224 |
def build_model():
|
| 225 |
+
if CEREBRAS_API_KEY:
|
| 226 |
+
m = pick_compat_model(
|
| 227 |
+
"https://api.cerebras.ai/v1/models", CEREBRAS_API_KEY,
|
| 228 |
+
("gpt-oss-120b", "llama-3.3-70b", "qwen-3-235b", "qwen", "llama"),
|
| 229 |
+
)
|
| 230 |
+
return LiteLLMModel(
|
| 231 |
+
model_id=f"cerebras/{m}", api_key=CEREBRAS_API_KEY, timeout=90, num_retries=4
|
| 232 |
+
)
|
| 233 |
+
if MISTRAL_API_KEY:
|
| 234 |
+
m = pick_compat_model(
|
| 235 |
+
"https://api.mistral.ai/v1/models", MISTRAL_API_KEY,
|
| 236 |
+
("mistral-large-latest", "mistral-large", "mistral-medium-latest", "mistral-medium", "mistral-small-latest", "mistral-small"),
|
| 237 |
+
)
|
| 238 |
+
return LiteLLMModel(
|
| 239 |
+
model_id=f"mistral/{m}", api_key=MISTRAL_API_KEY, timeout=90, num_retries=4
|
| 240 |
+
)
|
| 241 |
if GROQ_API_KEY:
|
| 242 |
return LiteLLMModel(
|
| 243 |
model_id=f"groq/{pick_groq_model()}",
|
|
|
|
| 254 |
)
|
| 255 |
if GEMINI_API_KEY:
|
| 256 |
return LiteLLMModel(
|
| 257 |
+
model_id=f"gemini/{gemini_model_name()}",
|
| 258 |
api_key=GEMINI_API_KEY,
|
| 259 |
+
timeout=120,
|
| 260 |
+
num_retries=6,
|
| 261 |
)
|
|
|
|
| 262 |
return InferenceClientModel(
|
| 263 |
model_id=os.getenv("MODEL_ID", "Qwen/Qwen2.5-72B-Instruct"), token=HF_TOKEN
|
| 264 |
)
|
|
|
|
| 279 |
self.agent = CodeAgent(
|
| 280 |
tools=[
|
| 281 |
DuckDuckGoSearchTool(),
|
| 282 |
+
VisitWebpageTool(max_output_length=6000),
|
| 283 |
download_task_file,
|
| 284 |
read_file,
|
| 285 |
transcribe_audio,
|
|
|
|
| 352 |
print(f"Task {task_id} failed: {err}")
|
| 353 |
answers.append({"task_id": task_id, "submitted_answer": ans})
|
| 354 |
log.append({"Task ID": task_id, "Question": question[:80], "Submitted Answer": ans, "Error": err})
|
| 355 |
+
time.sleep(12) # stay under free-tier rate limits
|
| 356 |
|
| 357 |
if not any(a["submitted_answer"] for a in answers):
|
| 358 |
first_err = next((l["Error"] for l in log if l["Error"]), "unknown")
|
|
|
|
| 386 |
try:
|
| 387 |
agent = BasicAgent(max_steps=3)
|
| 388 |
q = ".rewsna eht sa \"tfel\" drow eht fo etisoppo eht etirw ,ecnetnes siht dnatsrednu uoy fI"
|
| 389 |
+
used = agent.agent.model.model_id
|
| 390 |
return f"Model: {used}\nAnswer: {agent(q, 'test', '')} (expected: right)"
|
| 391 |
except Exception as e:
|
| 392 |
return f"FAILED: {type(e).__name__}: {e}"
|