Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -22,7 +22,6 @@ Pravidla pro tebe:
|
|
| 22 |
4. Pamatuj si, co uživatel říkal v předchozích větách této konverzace.
|
| 23 |
""".strip()
|
| 24 |
|
| 25 |
-
|
| 26 |
st.set_page_config(page_title="AI Rádce s pamětí", layout="centered")
|
| 27 |
st.title("🧠 Chytrý Chatbot (s pamětí)")
|
| 28 |
|
|
@@ -32,35 +31,16 @@ MODEL_REPO = "QuantFactory/Meta-Llama-3-8B-Instruct-GGUF"
|
|
| 32 |
MODEL_FILE = "Meta-Llama-3-8B-Instruct.Q4_K_M.gguf"
|
| 33 |
|
| 34 |
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
urls_text = st.text_area(
|
| 39 |
-
"URL zdroje (1 URL na řádek)",
|
| 40 |
-
value="\n".join(DEFAULT_URLS),
|
| 41 |
-
height=110,
|
| 42 |
-
)
|
| 43 |
-
urls = [u.strip() for u in urls_text.splitlines() if u.strip()]
|
| 44 |
-
|
| 45 |
-
max_new_tokens = st.slider("Max nových tokenů (rychlost)", 32, 256, 128, 16)
|
| 46 |
-
context_window = st.select_slider("Context window", options=[1024, 2048, 3072, 4096], value=2048)
|
| 47 |
-
|
| 48 |
-
cpu_cnt = os.cpu_count() or 2
|
| 49 |
-
threads = st.slider("Počet vláken (threads)", 1, min(8, cpu_cnt), min(4, cpu_cnt), 1)
|
| 50 |
-
batch = st.select_slider("Batch", options=[64, 128, 256, 512], value=256)
|
| 51 |
-
|
| 52 |
-
if st.button("🧹 Resetovat konverzaci"):
|
| 53 |
-
st.session_state.pop("messages", None)
|
| 54 |
-
st.session_state.pop("chat_engine", None)
|
| 55 |
-
st.rerun()
|
| 56 |
|
| 57 |
|
| 58 |
def create_llm(model_path: str, ctx_win: int, max_tok: int, n_threads: int, n_batch: int) -> LlamaCPP:
|
| 59 |
"""
|
| 60 |
-
Kompatibilní konstrukce
|
| 61 |
-
|
| 62 |
"""
|
| 63 |
-
# 1)
|
| 64 |
try:
|
| 65 |
return LlamaCPP(
|
| 66 |
model_path=model_path,
|
|
@@ -74,7 +54,7 @@ def create_llm(model_path: str, ctx_win: int, max_tok: int, n_threads: int, n_ba
|
|
| 74 |
except TypeError:
|
| 75 |
pass
|
| 76 |
|
| 77 |
-
# 2) fallback přes model_kwargs
|
| 78 |
try:
|
| 79 |
return LlamaCPP(
|
| 80 |
model_path=model_path,
|
|
@@ -85,7 +65,7 @@ def create_llm(model_path: str, ctx_win: int, max_tok: int, n_threads: int, n_ba
|
|
| 85 |
verbose=False,
|
| 86 |
)
|
| 87 |
except TypeError:
|
| 88 |
-
# 3) poslední fallback
|
| 89 |
return LlamaCPP(
|
| 90 |
model_path=model_path,
|
| 91 |
temperature=0.1,
|
|
@@ -97,42 +77,97 @@ def create_llm(model_path: str, ctx_win: int, max_tok: int, n_threads: int, n_ba
|
|
| 97 |
|
| 98 |
|
| 99 |
@st.cache_resource
|
| 100 |
-
def
|
| 101 |
-
|
| 102 |
-
model_path = hf_hub_download(repo_id=MODEL_REPO, filename=MODEL_FILE)
|
| 103 |
|
| 104 |
-
llm = create_llm(model_path, ctx_win, max_tok, n_threads, n_batch)
|
| 105 |
|
| 106 |
-
|
|
|
|
|
|
|
| 107 |
Settings.embed_model = FastEmbedEmbedding(model_name="BAAI/bge-small-en-v1.5")
|
| 108 |
|
| 109 |
docs = SimpleWebPageReader(html_to_text=True).load_data(list(urls_tuple))
|
| 110 |
return VectorStoreIndex.from_documents(docs)
|
| 111 |
|
| 112 |
|
| 113 |
-
|
| 114 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 115 |
|
| 116 |
-
|
| 117 |
-
|
| 118 |
|
|
|
|
|
|
|
|
|
|
| 119 |
return index.as_chat_engine(
|
| 120 |
-
chat_mode="
|
| 121 |
memory=memory,
|
| 122 |
system_prompt=SYSTEM_PROMPT,
|
| 123 |
verbose=False,
|
| 124 |
)
|
| 125 |
|
| 126 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 127 |
if "chat_engine" not in st.session_state:
|
| 128 |
-
with st.spinner("Startuji mozek bota... (
|
| 129 |
try:
|
| 130 |
-
st.session_state.chat_engine = make_chat_engine()
|
| 131 |
except Exception as e:
|
| 132 |
st.error(f"Chyba při inicializaci: {e}")
|
| 133 |
st.stop()
|
| 134 |
|
| 135 |
-
|
| 136 |
if "messages" not in st.session_state:
|
| 137 |
st.session_state.messages = []
|
| 138 |
|
|
@@ -141,6 +176,47 @@ for msg in st.session_state.messages:
|
|
| 141 |
st.markdown(msg["content"])
|
| 142 |
|
| 143 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 144 |
prompt = st.chat_input("Zeptej se (např: Co umíš?)...")
|
| 145 |
if prompt:
|
| 146 |
st.session_state.messages.append({"role": "user", "content": prompt})
|
|
@@ -149,27 +225,25 @@ if prompt:
|
|
| 149 |
|
| 150 |
with st.chat_message("assistant"):
|
| 151 |
placeholder = st.empty()
|
| 152 |
-
|
| 153 |
-
t0 = time.time()
|
| 154 |
|
| 155 |
-
|
| 156 |
-
|
| 157 |
-
stream = st.session_state.chat_engine.stream_chat(prompt)
|
| 158 |
-
for chunk in stream.response_gen:
|
| 159 |
-
full += chunk
|
| 160 |
-
placeholder.markdown(full)
|
| 161 |
|
| 162 |
-
|
| 163 |
-
|
| 164 |
-
|
|
|
|
|
|
|
|
|
|
| 165 |
|
| 166 |
-
|
| 167 |
-
full = f"Chyba při generování odpovědi: {e}"
|
| 168 |
-
placeholder.markdown(full)
|
| 169 |
|
| 170 |
-
|
|
|
|
|
|
|
| 171 |
|
| 172 |
-
st.session_state.messages.append({"role": "assistant", "content":
|
| 173 |
|
| 174 |
|
| 175 |
|
|
|
|
| 22 |
4. Pamatuj si, co uživatel říkal v předchozích větách této konverzace.
|
| 23 |
""".strip()
|
| 24 |
|
|
|
|
| 25 |
st.set_page_config(page_title="AI Rádce s pamětí", layout="centered")
|
| 26 |
st.title("🧠 Chytrý Chatbot (s pamětí)")
|
| 27 |
|
|
|
|
| 31 |
MODEL_FILE = "Meta-Llama-3-8B-Instruct.Q4_K_M.gguf"
|
| 32 |
|
| 33 |
|
| 34 |
+
def _clamp(v: int, lo: int, hi: int) -> int:
|
| 35 |
+
return max(lo, min(hi, v))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 36 |
|
| 37 |
|
| 38 |
def create_llm(model_path: str, ctx_win: int, max_tok: int, n_threads: int, n_batch: int) -> LlamaCPP:
|
| 39 |
"""
|
| 40 |
+
Kompatibilní konstrukce napříč verzemi llama-index.
|
| 41 |
+
Někdy wrapper nepřijme n_threads/n_batch přímo => použijeme model_kwargs.
|
| 42 |
"""
|
| 43 |
+
# 1) zkus přímé parametry
|
| 44 |
try:
|
| 45 |
return LlamaCPP(
|
| 46 |
model_path=model_path,
|
|
|
|
| 54 |
except TypeError:
|
| 55 |
pass
|
| 56 |
|
| 57 |
+
# 2) fallback přes model_kwargs
|
| 58 |
try:
|
| 59 |
return LlamaCPP(
|
| 60 |
model_path=model_path,
|
|
|
|
| 65 |
verbose=False,
|
| 66 |
)
|
| 67 |
except TypeError:
|
| 68 |
+
# 3) poslední fallback: jen threads
|
| 69 |
return LlamaCPP(
|
| 70 |
model_path=model_path,
|
| 71 |
temperature=0.1,
|
|
|
|
| 77 |
|
| 78 |
|
| 79 |
@st.cache_resource
|
| 80 |
+
def get_model_path() -> str:
|
| 81 |
+
return hf_hub_download(repo_id=MODEL_REPO, filename=MODEL_FILE)
|
|
|
|
| 82 |
|
|
|
|
| 83 |
|
| 84 |
+
@st.cache_resource
|
| 85 |
+
def build_index(urls_tuple: tuple[str, ...]) -> VectorStoreIndex:
|
| 86 |
+
# Embed model: bez torch/cuda (rychlejší instalace a stabilní na HF)
|
| 87 |
Settings.embed_model = FastEmbedEmbedding(model_name="BAAI/bge-small-en-v1.5")
|
| 88 |
|
| 89 |
docs = SimpleWebPageReader(html_to_text=True).load_data(list(urls_tuple))
|
| 90 |
return VectorStoreIndex.from_documents(docs)
|
| 91 |
|
| 92 |
|
| 93 |
+
@st.cache_resource
|
| 94 |
+
def load_llm_cached(ctx_win: int, max_tok: int, n_threads: int, n_batch: int) -> LlamaCPP:
|
| 95 |
+
model_path = get_model_path()
|
| 96 |
+
return create_llm(model_path, ctx_win, max_tok, n_threads, n_batch)
|
| 97 |
+
|
| 98 |
+
|
| 99 |
+
def make_chat_engine(urls_list: list[str], ctx_win: int, max_tok: int, n_threads: int, n_batch: int) -> object:
|
| 100 |
+
index = build_index(tuple(urls_list))
|
| 101 |
|
| 102 |
+
llm = load_llm_cached(ctx_win, max_tok, n_threads, n_batch)
|
| 103 |
+
Settings.llm = llm # nastavíme aktivní LLM pro LlamaIndex
|
| 104 |
|
| 105 |
+
memory = ChatMemoryBuffer.from_defaults(token_limit=min(1500, ctx_win))
|
| 106 |
+
|
| 107 |
+
# condense_plus_context bývá svižnější / stabilnější než čisté "context"
|
| 108 |
return index.as_chat_engine(
|
| 109 |
+
chat_mode="condense_plus_context",
|
| 110 |
memory=memory,
|
| 111 |
system_prompt=SYSTEM_PROMPT,
|
| 112 |
verbose=False,
|
| 113 |
)
|
| 114 |
|
| 115 |
|
| 116 |
+
with st.sidebar:
|
| 117 |
+
st.header("Nastavení")
|
| 118 |
+
|
| 119 |
+
urls_text = st.text_area(
|
| 120 |
+
"URL zdroje (1 URL na řádek)",
|
| 121 |
+
value="\n".join(DEFAULT_URLS),
|
| 122 |
+
height=110,
|
| 123 |
+
)
|
| 124 |
+
urls = [u.strip() for u in urls_text.splitlines() if u.strip()]
|
| 125 |
+
|
| 126 |
+
safe_mode = st.toggle("Safe Mode (doporučeno pro HF CPU)", value=True)
|
| 127 |
+
st.caption("Safe Mode brání nastavení, které na HF CPU typicky 'zamrzne'.")
|
| 128 |
+
|
| 129 |
+
# Uživatelské vstupy
|
| 130 |
+
user_max_new_tokens = st.slider("Max nových tokenů (rychlost)", 32, 256, 96, 16)
|
| 131 |
+
user_context_window = st.select_slider("Context window", options=[1024, 2048, 3072, 4096], value=2048)
|
| 132 |
+
|
| 133 |
+
cpu_cnt = os.cpu_count() or 2
|
| 134 |
+
user_threads = st.slider("Počet vláken (threads)", 1, min(8, cpu_cnt), min(4, cpu_cnt), 1)
|
| 135 |
+
user_batch = st.select_slider("Batch", options=[64, 128, 256, 512], value=128)
|
| 136 |
+
|
| 137 |
+
if st.button("🧹 Resetovat konverzaci"):
|
| 138 |
+
st.session_state.pop("messages", None)
|
| 139 |
+
st.session_state.pop("chat_engine", None)
|
| 140 |
+
st.rerun()
|
| 141 |
+
|
| 142 |
+
# Tvrdé limity (aby se to nezabilo na HF CPU)
|
| 143 |
+
if safe_mode:
|
| 144 |
+
max_new_tokens = _clamp(user_max_new_tokens, 32, 128)
|
| 145 |
+
context_window = _clamp(user_context_window, 1024, 2048)
|
| 146 |
+
threads = _clamp(user_threads, 1, 4)
|
| 147 |
+
batch = _clamp(user_batch, 64, 256)
|
| 148 |
+
else:
|
| 149 |
+
max_new_tokens = user_max_new_tokens
|
| 150 |
+
context_window = user_context_window
|
| 151 |
+
threads = user_threads
|
| 152 |
+
batch = user_batch
|
| 153 |
+
|
| 154 |
+
st.sidebar.markdown("---")
|
| 155 |
+
st.sidebar.write("Aktivní parametry:")
|
| 156 |
+
st.sidebar.code(
|
| 157 |
+
f"max_new_tokens={max_new_tokens}\ncontext_window={context_window}\nthreads={threads}\nbatch={batch}"
|
| 158 |
+
)
|
| 159 |
+
|
| 160 |
+
|
| 161 |
+
# Inicializace enginu
|
| 162 |
if "chat_engine" not in st.session_state:
|
| 163 |
+
with st.spinner("Startuji mozek bota... (model + index)"):
|
| 164 |
try:
|
| 165 |
+
st.session_state.chat_engine = make_chat_engine(urls, context_window, max_new_tokens, threads, batch)
|
| 166 |
except Exception as e:
|
| 167 |
st.error(f"Chyba při inicializaci: {e}")
|
| 168 |
st.stop()
|
| 169 |
|
| 170 |
+
# Historie zpráv
|
| 171 |
if "messages" not in st.session_state:
|
| 172 |
st.session_state.messages = []
|
| 173 |
|
|
|
|
| 176 |
st.markdown(msg["content"])
|
| 177 |
|
| 178 |
|
| 179 |
+
def generate_answer(prompt: str) -> str:
|
| 180 |
+
"""
|
| 181 |
+
Robustní generace odpovědi:
|
| 182 |
+
- zkusíme stream_chat (když funguje)
|
| 183 |
+
- pokud do 3s nepřiteče žádný chunk, fallback na chat()
|
| 184 |
+
"""
|
| 185 |
+
engine = st.session_state.chat_engine
|
| 186 |
+
|
| 187 |
+
# 1) Stream pokus
|
| 188 |
+
try:
|
| 189 |
+
stream = engine.stream_chat(prompt)
|
| 190 |
+
full = ""
|
| 191 |
+
started = time.time()
|
| 192 |
+
|
| 193 |
+
# Některé verze blokují; proto "čekáme na první chunk" max 3s
|
| 194 |
+
got_any = False
|
| 195 |
+
for chunk in stream.response_gen:
|
| 196 |
+
got_any = True
|
| 197 |
+
full += chunk
|
| 198 |
+
yield ("stream", full) # průběžně vracíme text
|
| 199 |
+
# když se to rozjede, necháme to dojet normálně
|
| 200 |
+
|
| 201 |
+
if got_any and full.strip():
|
| 202 |
+
return # výstup už byl odeslán přes yield
|
| 203 |
+
|
| 204 |
+
# pokud nic nepřišlo, padneme do fallbacku
|
| 205 |
+
if time.time() - started < 3.0:
|
| 206 |
+
# malá pauza, ať se neflushuje zbytečně
|
| 207 |
+
time.sleep(0.2)
|
| 208 |
+
|
| 209 |
+
except Exception:
|
| 210 |
+
# stream nemusí být podporovaný/kompatibilní
|
| 211 |
+
pass
|
| 212 |
+
|
| 213 |
+
# 2) Fallback: klasický chat() (blokuje, ale aspoň funguje vždy)
|
| 214 |
+
resp = engine.chat(prompt)
|
| 215 |
+
answer = getattr(resp, "response", None) or str(resp)
|
| 216 |
+
yield ("final", answer)
|
| 217 |
+
|
| 218 |
+
|
| 219 |
+
# Chat input
|
| 220 |
prompt = st.chat_input("Zeptej se (např: Co umíš?)...")
|
| 221 |
if prompt:
|
| 222 |
st.session_state.messages.append({"role": "user", "content": prompt})
|
|
|
|
| 225 |
|
| 226 |
with st.chat_message("assistant"):
|
| 227 |
placeholder = st.empty()
|
| 228 |
+
status = st.empty()
|
|
|
|
| 229 |
|
| 230 |
+
full_text = ""
|
| 231 |
+
t0 = time.time()
|
|
|
|
|
|
|
|
|
|
|
|
|
| 232 |
|
| 233 |
+
try:
|
| 234 |
+
# průběžné vykreslování
|
| 235 |
+
for kind, text in generate_answer(prompt):
|
| 236 |
+
full_text = text
|
| 237 |
+
placeholder.markdown(full_text)
|
| 238 |
+
status.caption(f"Generuji... {time.time() - t0:.1f}s")
|
| 239 |
|
| 240 |
+
status.caption(f"Hotovo za {time.time() - t0:.1f}s")
|
|
|
|
|
|
|
| 241 |
|
| 242 |
+
except Exception as e:
|
| 243 |
+
full_text = f"Chyba při generování odpovědi: {e}"
|
| 244 |
+
placeholder.markdown(full_text)
|
| 245 |
|
| 246 |
+
st.session_state.messages.append({"role": "assistant", "content": full_text})
|
| 247 |
|
| 248 |
|
| 249 |
|