Spaces:
Sleeping
Sleeping
Commit ·
24766ca
1
Parent(s): dac3615
add openai data export compatibility
Browse files- app.py +62 -68
- requirements.txt +0 -1
app.py
CHANGED
|
@@ -1,14 +1,14 @@
|
|
| 1 |
-
# app.py -
|
| 2 |
-
# Beschreibung:
|
| 3 |
-
#
|
| 4 |
-
#
|
|
|
|
| 5 |
|
| 6 |
import os
|
| 7 |
import torch
|
| 8 |
import gradio as gr
|
| 9 |
import time
|
| 10 |
-
import
|
| 11 |
-
import ijson # <-- DAS RICHTIGE WERKZEUG
|
| 12 |
|
| 13 |
from typing import List, Tuple, Generator, Dict
|
| 14 |
from threading import Thread
|
|
@@ -72,70 +72,38 @@ def get_llm() -> Tuple[Gemma3ForConditionalGeneration, AutoProcessor]:
|
|
| 72 |
# --------------------------------------------------------------------
|
| 73 |
# Datei-Handling & Chunking
|
| 74 |
# --------------------------------------------------------------------
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
"""
|
| 78 |
-
Verwendet `ijson`, um eine JSON-Datei iterativ zu parsen. Dies ist extrem robust
|
| 79 |
-
gegenüber Dateigröße und strukturellen Fehlern wie einer fehlenden schließenden Klammer.
|
| 80 |
-
"""
|
| 81 |
-
print_debug(f"Starte professionelles Parsen (ijson) der Datei: {file_path}")
|
| 82 |
-
all_conversations_text = []
|
| 83 |
|
| 84 |
try:
|
| 85 |
-
with open(file_path, '
|
| 86 |
-
|
| 87 |
-
parser = ijson.items(f, 'item')
|
| 88 |
-
for i, conversation in enumerate(parser):
|
| 89 |
-
try:
|
| 90 |
-
conversation_parts = []
|
| 91 |
-
title = conversation.get('title', f'Unbenannte Konversation {i+1}')
|
| 92 |
-
mapping = conversation.get('mapping')
|
| 93 |
-
|
| 94 |
-
if not mapping or not isinstance(mapping, dict): continue
|
| 95 |
-
|
| 96 |
-
for node_id, node in mapping.items():
|
| 97 |
-
message = node.get('message')
|
| 98 |
-
if not message: continue
|
| 99 |
-
try:
|
| 100 |
-
role = message['author']['role']
|
| 101 |
-
content_node = message.get('content', {})
|
| 102 |
-
parts = content_node.get('parts', [])
|
| 103 |
-
|
| 104 |
-
if role in ['user', 'assistant'] and parts and isinstance(parts, list) and parts[0]:
|
| 105 |
-
text_content = ""
|
| 106 |
-
if isinstance(parts[0], str):
|
| 107 |
-
text_content = parts[0]
|
| 108 |
-
elif isinstance(parts[0], dict) and 'text' in parts[0]:
|
| 109 |
-
text_content = parts[0].get('text', "")
|
| 110 |
-
|
| 111 |
-
if text_content and text_content.strip():
|
| 112 |
-
conversation_parts.append(f"{role.capitalize()}: {text_content.strip()}")
|
| 113 |
-
except (KeyError, TypeError, IndexError):
|
| 114 |
-
continue
|
| 115 |
-
|
| 116 |
-
if conversation_parts:
|
| 117 |
-
full_conversation = "\n\n".join(conversation_parts)
|
| 118 |
-
all_conversations_text.append(f"--- Konversation Start: {title} ---\n\n{full_conversation}\n\n--- Konversation Ende: {title} ---")
|
| 119 |
-
print_debug(f"Konversation '{title}' erfolgreich via ijson extrahiert.")
|
| 120 |
-
|
| 121 |
-
except Exception as e:
|
| 122 |
-
print_debug(f"Unerwarteter Fehler bei der Verarbeitung einer Konversation von ijson. Überspringe. Fehler: {e}")
|
| 123 |
-
continue
|
| 124 |
-
except ijson.JSONError as e:
|
| 125 |
-
print_debug(f"Schwerwiegender JSON-Fehler, ijson konnte nicht parsen: {e}")
|
| 126 |
-
# Selbst wenn ein schwerer Fehler auftritt, haben wir vielleicht schon Konversationen gesammelt
|
| 127 |
except Exception as e:
|
| 128 |
-
print_debug(f"Konnte Datei nicht lesen
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 129 |
return ""
|
| 130 |
|
| 131 |
-
|
| 132 |
-
|
|
|
|
|
|
|
|
|
|
| 133 |
return final_text
|
| 134 |
|
| 135 |
def extract_text_from_file(path: str) -> str:
|
| 136 |
ext = os.path.splitext(path)[1].lower()
|
| 137 |
if ext == ".json":
|
| 138 |
-
return
|
| 139 |
if ext in [".txt", ".md", ".markdown"]:
|
| 140 |
with open(path, "r", encoding="utf-8", errors="ignore") as f: return f.read()
|
| 141 |
if ext == ".pdf":
|
|
@@ -152,9 +120,14 @@ def extract_text_from_file(path: str) -> str:
|
|
| 152 |
with open(path, "r", encoding="utf-8", errors="ignore") as f: return f.read()
|
| 153 |
except Exception: return ""
|
| 154 |
|
| 155 |
-
# ... [Rest des Codes für Indexing, RAG und UI bleibt identisch] ...
|
| 156 |
def get_text_splitter() -> RecursiveCharacterTextSplitter:
|
| 157 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 158 |
|
| 159 |
def index_files(file_paths: List[str], progress=gr.Progress(track_tqdm=True)) -> str:
|
| 160 |
global VECTOR_STORE
|
|
@@ -195,12 +168,23 @@ def clear_index() -> str:
|
|
| 195 |
return "Index geleert."
|
| 196 |
|
| 197 |
def retrieve_relevant_chunks(query: str, top_k: int = 5) -> List[Dict]:
|
| 198 |
-
if VECTOR_STORE is None:
|
|
|
|
|
|
|
|
|
|
|
|
|
| 199 |
results_with_scores = VECTOR_STORE.similarity_search_with_score(query, k=top_k)
|
| 200 |
-
|
|
|
|
| 201 |
"content": doc.page_content, "source": doc.metadata.get("source", "Unbekannt"), "score": 1 - score
|
| 202 |
} for doc, score in results_with_scores]
|
| 203 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 204 |
def build_rag_prompt(user_question: str, retrieved_chunks: List[Dict]) -> str:
|
| 205 |
if not retrieved_chunks: context_str = "Es wurden keine relevanten Dokumente im Kontext gefunden."
|
| 206 |
else:
|
|
@@ -218,30 +202,41 @@ def build_rag_prompt(user_question: str, retrieved_chunks: List[Dict]) -> str:
|
|
| 218 |
return prompt
|
| 219 |
|
| 220 |
def answer_with_rag(question: str, history: list) -> Generator[str, None, None]:
|
|
|
|
| 221 |
model, processor = get_llm()
|
| 222 |
streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
|
|
|
|
| 223 |
retrieved = retrieve_relevant_chunks(question, top_k=5)
|
| 224 |
prompt = build_rag_prompt(question, retrieved)
|
|
|
|
|
|
|
|
|
|
| 225 |
messages = [{"role": "user", "content": [{"type": "text", "text": prompt}]}]
|
| 226 |
input_ids = processor.apply_chat_template(
|
| 227 |
messages, tokenize=True, add_generation_prompt=True, return_tensors="pt"
|
| 228 |
).to(model.device)
|
|
|
|
| 229 |
generation_kwargs = {
|
| 230 |
"input_ids": input_ids, "streamer": streamer, "max_new_tokens": 1024,
|
| 231 |
"do_sample": True, "temperature": 0.7, "top_p": 0.9,
|
| 232 |
}
|
| 233 |
thread = Thread(target=model.generate, kwargs=generation_kwargs)
|
| 234 |
thread.start()
|
|
|
|
|
|
|
| 235 |
for new_text in streamer:
|
|
|
|
| 236 |
yield new_text
|
| 237 |
|
|
|
|
|
|
|
| 238 |
def build_demo() -> gr.Blocks:
|
| 239 |
-
with gr.Blocks(title="Gemma 3 RAG
|
| 240 |
gr.Markdown(
|
| 241 |
"""
|
| 242 |
-
# 🔍 Gemma 3 RAG
|
| 243 |
**Analysiere deine (auch unvollständigen) ChatGPT `conversations.json` oder andere Dokumente.**
|
| 244 |
-
Verwendet
|
| 245 |
"""
|
| 246 |
)
|
| 247 |
with gr.Row():
|
|
@@ -260,7 +255,6 @@ def build_demo() -> gr.Blocks:
|
|
| 260 |
sample_path = "sample.json"
|
| 261 |
if os.path.exists(sample_path):
|
| 262 |
print_debug(f"'{sample_path}' gefunden. Starte Indexierung via Button.")
|
| 263 |
-
# Zeige dem User, dass etwas passiert
|
| 264 |
yield "Indexiere `sample.json`..."
|
| 265 |
status = index_files([sample_path], progress=gr.Progress(track_tqdm=True))
|
| 266 |
yield status
|
|
|
|
| 1 |
+
# app.py - v3.2 (The Final Production Version)
|
| 2 |
+
# Beschreibung: Die finale, korrigierte und produktionsreife Version.
|
| 3 |
+
# Behebt den kritischen Chunking-Fehler, indem auf eine natürliche
|
| 4 |
+
# Textformatierung und die Standard-Splitting-Strategie zurückgegriffen wird.
|
| 5 |
+
# Behält die volle Debug-Ausgabe für Transparenz.
|
| 6 |
|
| 7 |
import os
|
| 8 |
import torch
|
| 9 |
import gradio as gr
|
| 10 |
import time
|
| 11 |
+
import re
|
|
|
|
| 12 |
|
| 13 |
from typing import List, Tuple, Generator, Dict
|
| 14 |
from threading import Thread
|
|
|
|
| 72 |
# --------------------------------------------------------------------
|
| 73 |
# Datei-Handling & Chunking
|
| 74 |
# --------------------------------------------------------------------
|
| 75 |
+
def _regex_based_extractor(file_path: str) -> str:
|
| 76 |
+
print_debug(f"Starte Regex-basierten Parser (v3.2) für: {file_path}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 77 |
|
| 78 |
try:
|
| 79 |
+
with open(file_path, 'r', encoding='utf-8', errors='ignore') as f:
|
| 80 |
+
content = f.read()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 81 |
except Exception as e:
|
| 82 |
+
print_debug(f"Konnte Datei nicht lesen: {e}")
|
| 83 |
+
return ""
|
| 84 |
+
|
| 85 |
+
pattern = re.compile(
|
| 86 |
+
r'"role":\s*"(user|assistant)".*?"parts":\s*\[\s*"(.*?)"',
|
| 87 |
+
re.DOTALL
|
| 88 |
+
)
|
| 89 |
+
|
| 90 |
+
matches = pattern.findall(content)
|
| 91 |
+
|
| 92 |
+
if not matches:
|
| 93 |
+
print_debug("Keine passenden Textteile über Regex gefunden.")
|
| 94 |
return ""
|
| 95 |
|
| 96 |
+
# KORREKTUR: Wir erstellen einen natürlich formatierten Text.
|
| 97 |
+
full_text_parts = [f"{role.capitalize()}: {text.strip()}" for role, text in matches if text.strip()]
|
| 98 |
+
final_text = "\n\n".join(full_text_parts) # <-- EINFACHER, ROBUSTER SEPARATOR
|
| 99 |
+
|
| 100 |
+
print_debug(f"Gesamter Text aus JSON via Regex extrahiert ({len(final_text)} Zeichen aus {len(matches)} Treffern).")
|
| 101 |
return final_text
|
| 102 |
|
| 103 |
def extract_text_from_file(path: str) -> str:
|
| 104 |
ext = os.path.splitext(path)[1].lower()
|
| 105 |
if ext == ".json":
|
| 106 |
+
return _regex_based_extractor(path)
|
| 107 |
if ext in [".txt", ".md", ".markdown"]:
|
| 108 |
with open(path, "r", encoding="utf-8", errors="ignore") as f: return f.read()
|
| 109 |
if ext == ".pdf":
|
|
|
|
| 120 |
with open(path, "r", encoding="utf-8", errors="ignore") as f: return f.read()
|
| 121 |
except Exception: return ""
|
| 122 |
|
|
|
|
| 123 |
def get_text_splitter() -> RecursiveCharacterTextSplitter:
|
| 124 |
+
# KORREKTUR: Wir kehren zu den Standard-Separatoren zurück, die für
|
| 125 |
+
# natürlich formatierten Text optimiert sind.
|
| 126 |
+
return RecursiveCharacterTextSplitter(
|
| 127 |
+
chunk_size=1000,
|
| 128 |
+
chunk_overlap=200,
|
| 129 |
+
length_function=len
|
| 130 |
+
)
|
| 131 |
|
| 132 |
def index_files(file_paths: List[str], progress=gr.Progress(track_tqdm=True)) -> str:
|
| 133 |
global VECTOR_STORE
|
|
|
|
| 168 |
return "Index geleert."
|
| 169 |
|
| 170 |
def retrieve_relevant_chunks(query: str, top_k: int = 5) -> List[Dict]:
|
| 171 |
+
if VECTOR_STORE is None:
|
| 172 |
+
print_debug("Retrieval versucht, aber Vektor-Index ist leer.")
|
| 173 |
+
return []
|
| 174 |
+
|
| 175 |
+
print_debug(f"Suche nach {top_k} relevanten Chunks für die Anfrage: '{query}'")
|
| 176 |
results_with_scores = VECTOR_STORE.similarity_search_with_score(query, k=top_k)
|
| 177 |
+
|
| 178 |
+
formatted_results = [{
|
| 179 |
"content": doc.page_content, "source": doc.metadata.get("source", "Unbekannt"), "score": 1 - score
|
| 180 |
} for doc, score in results_with_scores]
|
| 181 |
|
| 182 |
+
print_debug(f"{len(formatted_results)} Chunks gefunden. Details:")
|
| 183 |
+
for i, chunk in enumerate(formatted_results):
|
| 184 |
+
print_debug(f" [Chunk {i+1}] Score: {chunk['score']:.4f}, Source: {chunk['source']}\n Content: '{chunk['content'][:200].replace(chr(10), ' ')}...'")
|
| 185 |
+
|
| 186 |
+
return formatted_results
|
| 187 |
+
|
| 188 |
def build_rag_prompt(user_question: str, retrieved_chunks: List[Dict]) -> str:
|
| 189 |
if not retrieved_chunks: context_str = "Es wurden keine relevanten Dokumente im Kontext gefunden."
|
| 190 |
else:
|
|
|
|
| 202 |
return prompt
|
| 203 |
|
| 204 |
def answer_with_rag(question: str, history: list) -> Generator[str, None, None]:
|
| 205 |
+
print_debug(f"--- RAG Pipeline Start für Frage: '{question}' ---")
|
| 206 |
model, processor = get_llm()
|
| 207 |
streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
|
| 208 |
+
|
| 209 |
retrieved = retrieve_relevant_chunks(question, top_k=5)
|
| 210 |
prompt = build_rag_prompt(question, retrieved)
|
| 211 |
+
|
| 212 |
+
print_debug(f"--- Vollständiger RAG-Prompt, der an das LLM gesendet wird ---\n{prompt}\n--- Ende des RAG-Prompts ---")
|
| 213 |
+
|
| 214 |
messages = [{"role": "user", "content": [{"type": "text", "text": prompt}]}]
|
| 215 |
input_ids = processor.apply_chat_template(
|
| 216 |
messages, tokenize=True, add_generation_prompt=True, return_tensors="pt"
|
| 217 |
).to(model.device)
|
| 218 |
+
|
| 219 |
generation_kwargs = {
|
| 220 |
"input_ids": input_ids, "streamer": streamer, "max_new_tokens": 1024,
|
| 221 |
"do_sample": True, "temperature": 0.7, "top_p": 0.9,
|
| 222 |
}
|
| 223 |
thread = Thread(target=model.generate, kwargs=generation_kwargs)
|
| 224 |
thread.start()
|
| 225 |
+
|
| 226 |
+
full_response = ""
|
| 227 |
for new_text in streamer:
|
| 228 |
+
full_response += new_text
|
| 229 |
yield new_text
|
| 230 |
|
| 231 |
+
print_debug(f"--- Vollständige Antwort vom LLM (gestreamt) ---\n{full_response}\n--- Ende der LLM-Antwort ---")
|
| 232 |
+
|
| 233 |
def build_demo() -> gr.Blocks:
|
| 234 |
+
with gr.Blocks(title="Gemma 3 RAG v3.2", theme="soft") as demo:
|
| 235 |
gr.Markdown(
|
| 236 |
"""
|
| 237 |
+
# 🔍 Gemma 3 RAG v3.2 – The Final Production Version
|
| 238 |
**Analysiere deine (auch unvollständigen) ChatGPT `conversations.json` oder andere Dokumente.**
|
| 239 |
+
Verwendet einen robusten Regex-Extraktor und korrektes Chunking für maximale Zuverlässigkeit.
|
| 240 |
"""
|
| 241 |
)
|
| 242 |
with gr.Row():
|
|
|
|
| 255 |
sample_path = "sample.json"
|
| 256 |
if os.path.exists(sample_path):
|
| 257 |
print_debug(f"'{sample_path}' gefunden. Starte Indexierung via Button.")
|
|
|
|
| 258 |
yield "Indexiere `sample.json`..."
|
| 259 |
status = index_files([sample_path], progress=gr.Progress(track_tqdm=True))
|
| 260 |
yield status
|
requirements.txt
CHANGED
|
@@ -18,4 +18,3 @@ langchain-text-splitters>=0.2.0
|
|
| 18 |
langchain-community>=0.2.0
|
| 19 |
langchain-huggingface>=0.0.3 # <-- NEU: Dediziertes Paket für HF-Integration
|
| 20 |
faiss-cpu>=1.8.0
|
| 21 |
-
ijson>=3.2.3
|
|
|
|
| 18 |
langchain-community>=0.2.0
|
| 19 |
langchain-huggingface>=0.0.3 # <-- NEU: Dediziertes Paket für HF-Integration
|
| 20 |
faiss-cpu>=1.8.0
|
|
|