Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -10,7 +10,6 @@ from llama_index.readers.web import SimpleWebPageReader
|
|
| 10 |
from llama_index.embeddings.fastembed import FastEmbedEmbedding
|
| 11 |
|
| 12 |
|
| 13 |
-
# --- KONFIGURACE OSOBNOSTI BOTA ---
|
| 14 |
SYSTEM_PROMPT = """
|
| 15 |
Jsi inteligentní český asistent, který pomáhá uživatelům hledat informace na zadaném webu.
|
| 16 |
Tvé jméno je AI Rádce.
|
|
@@ -23,20 +22,19 @@ Pravidla pro tebe:
|
|
| 23 |
4. Pamatuj si, co uživatel říkal v předchozích větách této konverzace.
|
| 24 |
""".strip()
|
| 25 |
|
|
|
|
| 26 |
st.set_page_config(page_title="AI Rádce s pamětí", layout="centered")
|
| 27 |
st.title("🧠 Chytrý Chatbot (s pamětí)")
|
| 28 |
|
| 29 |
-
# --- NASTAVENÍ ZDROJE DAT ---
|
| 30 |
DEFAULT_URLS = ["https://cs.wikipedia.org/wiki/Umělá_inteligence"]
|
| 31 |
|
| 32 |
-
# --- NASTAVENÍ MODELU (GGUF) ---
|
| 33 |
MODEL_REPO = "QuantFactory/Meta-Llama-3-8B-Instruct-GGUF"
|
| 34 |
MODEL_FILE = "Meta-Llama-3-8B-Instruct.Q4_K_M.gguf"
|
| 35 |
|
| 36 |
|
| 37 |
-
# Sidebar: konfigurace
|
| 38 |
with st.sidebar:
|
| 39 |
st.header("Nastavení")
|
|
|
|
| 40 |
urls_text = st.text_area(
|
| 41 |
"URL zdroje (1 URL na řádek)",
|
| 42 |
value="\n".join(DEFAULT_URLS),
|
|
@@ -47,10 +45,9 @@ with st.sidebar:
|
|
| 47 |
max_new_tokens = st.slider("Max nových tokenů (rychlost)", 32, 256, 128, 16)
|
| 48 |
context_window = st.select_slider("Context window", options=[1024, 2048, 3072, 4096], value=2048)
|
| 49 |
|
| 50 |
-
# Výkon: HF CPU typicky 2–4 jádra; víc často nepomůže
|
| 51 |
cpu_cnt = os.cpu_count() or 2
|
| 52 |
-
|
| 53 |
-
|
| 54 |
|
| 55 |
if st.button("🧹 Resetovat konverzaci"):
|
| 56 |
st.session_state.pop("messages", None)
|
|
@@ -58,39 +55,65 @@ with st.sidebar:
|
|
| 58 |
st.rerun()
|
| 59 |
|
| 60 |
|
| 61 |
-
|
| 62 |
-
def load_index_and_llm(urls_tuple: tuple[str, ...], ctx_win: int, max_tok: int, threads: int, batch: int) -> VectorStoreIndex:
|
| 63 |
"""
|
| 64 |
-
|
| 65 |
-
|
| 66 |
"""
|
| 67 |
-
# 1)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 68 |
model_path = hf_hub_download(repo_id=MODEL_REPO, filename=MODEL_FILE)
|
| 69 |
|
| 70 |
-
|
| 71 |
-
llm = LlamaCPP(
|
| 72 |
-
model_path=model_path,
|
| 73 |
-
temperature=0.1,
|
| 74 |
-
max_new_tokens=max_tok,
|
| 75 |
-
context_window=ctx_win,
|
| 76 |
-
n_threads=threads,
|
| 77 |
-
n_batch=batch,
|
| 78 |
-
verbose=False,
|
| 79 |
-
)
|
| 80 |
|
| 81 |
-
# 3) Nastavení LlamaIndex
|
| 82 |
Settings.llm = llm
|
| 83 |
Settings.embed_model = FastEmbedEmbedding(model_name="BAAI/bge-small-en-v1.5")
|
| 84 |
|
| 85 |
-
# 4) Data z webu + index
|
| 86 |
docs = SimpleWebPageReader(html_to_text=True).load_data(list(urls_tuple))
|
| 87 |
-
|
| 88 |
-
return index
|
| 89 |
|
| 90 |
|
| 91 |
def make_chat_engine() -> object:
|
| 92 |
-
|
| 93 |
-
|
|
|
|
| 94 |
memory = ChatMemoryBuffer.from_defaults(token_limit=min(3000, context_window))
|
| 95 |
|
| 96 |
return index.as_chat_engine(
|
|
@@ -101,7 +124,6 @@ def make_chat_engine() -> object:
|
|
| 101 |
)
|
| 102 |
|
| 103 |
|
| 104 |
-
# Inicializace enginu
|
| 105 |
if "chat_engine" not in st.session_state:
|
| 106 |
with st.spinner("Startuji mozek bota... (načítám model a web)"):
|
| 107 |
try:
|
|
@@ -110,7 +132,7 @@ if "chat_engine" not in st.session_state:
|
|
| 110 |
st.error(f"Chyba při inicializaci: {e}")
|
| 111 |
st.stop()
|
| 112 |
|
| 113 |
-
|
| 114 |
if "messages" not in st.session_state:
|
| 115 |
st.session_state.messages = []
|
| 116 |
|
|
@@ -119,41 +141,35 @@ for msg in st.session_state.messages:
|
|
| 119 |
st.markdown(msg["content"])
|
| 120 |
|
| 121 |
|
| 122 |
-
# --- CHAT LOOP (se streamováním) ---
|
| 123 |
prompt = st.chat_input("Zeptej se (např: Co umíš?)...")
|
| 124 |
if prompt:
|
| 125 |
-
# user message
|
| 126 |
st.session_state.messages.append({"role": "user", "content": prompt})
|
| 127 |
with st.chat_message("user"):
|
| 128 |
st.markdown(prompt)
|
| 129 |
|
| 130 |
-
# assistant streaming response
|
| 131 |
with st.chat_message("assistant"):
|
| 132 |
placeholder = st.empty()
|
| 133 |
full = ""
|
| 134 |
-
|
| 135 |
|
| 136 |
with st.spinner("Přemýšlím..."):
|
| 137 |
try:
|
| 138 |
stream = st.session_state.chat_engine.stream_chat(prompt)
|
| 139 |
-
|
| 140 |
-
# stream.response_gen generuje text po částech
|
| 141 |
for chunk in stream.response_gen:
|
| 142 |
full += chunk
|
| 143 |
placeholder.markdown(full)
|
| 144 |
|
| 145 |
-
# kdyby stream nevrátil nic (edge-case), aspoň něco
|
| 146 |
if not full.strip():
|
| 147 |
-
full = getattr(stream, "response", None) or "
|
| 148 |
placeholder.markdown(full)
|
| 149 |
|
| 150 |
except Exception as e:
|
| 151 |
full = f"Chyba při generování odpovědi: {e}"
|
| 152 |
placeholder.markdown(full)
|
| 153 |
|
| 154 |
-
|
| 155 |
-
st.caption(f"Hotovo za {elapsed:.1f}s")
|
| 156 |
|
| 157 |
st.session_state.messages.append({"role": "assistant", "content": full})
|
| 158 |
|
| 159 |
|
|
|
|
|
|
| 10 |
from llama_index.embeddings.fastembed import FastEmbedEmbedding
|
| 11 |
|
| 12 |
|
|
|
|
| 13 |
SYSTEM_PROMPT = """
|
| 14 |
Jsi inteligentní český asistent, který pomáhá uživatelům hledat informace na zadaném webu.
|
| 15 |
Tvé jméno je AI Rádce.
|
|
|
|
| 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 |
|
|
|
|
| 29 |
DEFAULT_URLS = ["https://cs.wikipedia.org/wiki/Umělá_inteligence"]
|
| 30 |
|
|
|
|
| 31 |
MODEL_REPO = "QuantFactory/Meta-Llama-3-8B-Instruct-GGUF"
|
| 32 |
MODEL_FILE = "Meta-Llama-3-8B-Instruct.Q4_K_M.gguf"
|
| 33 |
|
| 34 |
|
|
|
|
| 35 |
with st.sidebar:
|
| 36 |
st.header("Nastavení")
|
| 37 |
+
|
| 38 |
urls_text = st.text_area(
|
| 39 |
"URL zdroje (1 URL na řádek)",
|
| 40 |
value="\n".join(DEFAULT_URLS),
|
|
|
|
| 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)
|
|
|
|
| 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 LlamaCPP napříč verzemi llama-index.
|
| 61 |
+
Některé verze nepřijímají n_threads/n_batch přímo, ale jen přes model_kwargs.
|
| 62 |
"""
|
| 63 |
+
# 1) zkusit přímé parametry (novější/verze dle wrapperu)
|
| 64 |
+
try:
|
| 65 |
+
return LlamaCPP(
|
| 66 |
+
model_path=model_path,
|
| 67 |
+
temperature=0.1,
|
| 68 |
+
max_new_tokens=max_tok,
|
| 69 |
+
context_window=ctx_win,
|
| 70 |
+
n_threads=n_threads,
|
| 71 |
+
n_batch=n_batch,
|
| 72 |
+
verbose=False,
|
| 73 |
+
)
|
| 74 |
+
except TypeError:
|
| 75 |
+
pass
|
| 76 |
+
|
| 77 |
+
# 2) fallback přes model_kwargs (časté u LlamaIndex wrapperu)
|
| 78 |
+
try:
|
| 79 |
+
return LlamaCPP(
|
| 80 |
+
model_path=model_path,
|
| 81 |
+
temperature=0.1,
|
| 82 |
+
max_new_tokens=max_tok,
|
| 83 |
+
context_window=ctx_win,
|
| 84 |
+
model_kwargs={"n_threads": n_threads, "n_batch": n_batch},
|
| 85 |
+
verbose=False,
|
| 86 |
+
)
|
| 87 |
+
except TypeError:
|
| 88 |
+
# 3) poslední fallback – jen threads (někdy n_batch není podporovaný)
|
| 89 |
+
return LlamaCPP(
|
| 90 |
+
model_path=model_path,
|
| 91 |
+
temperature=0.1,
|
| 92 |
+
max_new_tokens=max_tok,
|
| 93 |
+
context_window=ctx_win,
|
| 94 |
+
model_kwargs={"n_threads": n_threads},
|
| 95 |
+
verbose=False,
|
| 96 |
+
)
|
| 97 |
+
|
| 98 |
+
|
| 99 |
+
@st.cache_resource
|
| 100 |
+
def load_index_and_settings(urls_tuple: tuple[str, ...], ctx_win: int, max_tok: int, n_threads: int, n_batch: int) -> VectorStoreIndex:
|
| 101 |
+
# stáhnout GGUF do HF cache
|
| 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 |
Settings.llm = llm
|
| 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 |
def make_chat_engine() -> object:
|
| 114 |
+
index = load_index_and_settings(tuple(urls), context_window, max_new_tokens, threads, batch)
|
| 115 |
+
|
| 116 |
+
# paměť per-session (NEcacheovat)
|
| 117 |
memory = ChatMemoryBuffer.from_defaults(token_limit=min(3000, context_window))
|
| 118 |
|
| 119 |
return index.as_chat_engine(
|
|
|
|
| 124 |
)
|
| 125 |
|
| 126 |
|
|
|
|
| 127 |
if "chat_engine" not in st.session_state:
|
| 128 |
with st.spinner("Startuji mozek bota... (načítám model a web)"):
|
| 129 |
try:
|
|
|
|
| 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 |
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})
|
| 147 |
with st.chat_message("user"):
|
| 148 |
st.markdown(prompt)
|
| 149 |
|
|
|
|
| 150 |
with st.chat_message("assistant"):
|
| 151 |
placeholder = st.empty()
|
| 152 |
full = ""
|
| 153 |
+
t0 = time.time()
|
| 154 |
|
| 155 |
with st.spinner("Přemýšlím..."):
|
| 156 |
try:
|
| 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 |
if not full.strip():
|
| 163 |
+
full = getattr(stream, "response", None) or "Nedostal jsem žádná data k odpovědi."
|
| 164 |
placeholder.markdown(full)
|
| 165 |
|
| 166 |
except Exception as e:
|
| 167 |
full = f"Chyba při generování odpovědi: {e}"
|
| 168 |
placeholder.markdown(full)
|
| 169 |
|
| 170 |
+
st.caption(f"Hotovo za {time.time() - t0:.1f}s")
|
|
|
|
| 171 |
|
| 172 |
st.session_state.messages.append({"role": "assistant", "content": full})
|
| 173 |
|
| 174 |
|
| 175 |
+
|