Spaces:
Sleeping
Sleeping
Update tools_engine.py
Browse files- tools_engine.py +62 -35
tools_engine.py
CHANGED
|
@@ -1,74 +1,101 @@
|
|
| 1 |
"""
|
| 2 |
-
Nexari Tools Engine (
|
| 3 |
Author: Piyush
|
| 4 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
"""
|
| 6 |
|
| 7 |
from duckduckgo_search import DDGS
|
| 8 |
from transformers import pipeline
|
|
|
|
| 9 |
|
| 10 |
print(">>> Tools: Loading Intent Classification Model...")
|
|
|
|
| 11 |
intent_classifier = pipeline("zero-shot-classification", model="typeform/distilbert-base-uncased-mnli")
|
| 12 |
|
| 13 |
def analyze_intent(user_text):
|
| 14 |
"""
|
| 15 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 16 |
"""
|
|
|
|
|
|
|
|
|
|
| 17 |
text_lower = user_text.lower().strip()
|
| 18 |
-
|
| 19 |
-
# ===
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
|
|
|
| 24 |
if text_lower in direct_chat_triggers or any(text_lower.startswith(t + " ") for t in direct_chat_triggers):
|
| 25 |
-
print(f">>> Brain Override: Detected Greeting/Identity -> Force 'general
|
| 26 |
-
return "general
|
| 27 |
|
| 28 |
-
# ===
|
| 29 |
candidate_labels = [
|
| 30 |
-
"internet search",
|
| 31 |
-
"general conversation",
|
| 32 |
-
"coding request",
|
| 33 |
"checking time"
|
| 34 |
]
|
| 35 |
-
|
| 36 |
try:
|
| 37 |
result = intent_classifier(user_text, candidate_labels)
|
| 38 |
-
|
| 39 |
confidence = result['scores'][0]
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
# Confidence Threshold badha diya (0.4 -> 0.5) taaki ghalat search na kare
|
| 44 |
if confidence > 0.5:
|
| 45 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 46 |
except Exception as e:
|
| 47 |
print(f"Intent Error: {e}")
|
| 48 |
-
|
| 49 |
-
return "general conversation"
|
| 50 |
|
| 51 |
-
|
|
|
|
|
|
|
| 52 |
"""
|
| 53 |
-
|
| 54 |
"""
|
| 55 |
try:
|
| 56 |
clean_query = user_text.lower()
|
| 57 |
-
remove_phrases = ["search for", "google", "find", "tell me about", "latest info on", "news about"]
|
| 58 |
for phrase in remove_phrases:
|
| 59 |
clean_query = clean_query.replace(phrase, "")
|
| 60 |
-
|
| 61 |
clean_query = clean_query.strip()
|
| 62 |
-
if len(clean_query) < 2:
|
|
|
|
| 63 |
|
| 64 |
print(f">>> Action: Searching Web for '{clean_query}'...")
|
| 65 |
-
results = DDGS().text(clean_query, max_results=
|
| 66 |
-
|
| 67 |
if results:
|
| 68 |
-
|
| 69 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 70 |
return ""
|
| 71 |
-
|
| 72 |
except Exception as e:
|
| 73 |
print(f"Search Error: {e}")
|
| 74 |
-
return ""
|
|
|
|
| 1 |
"""
|
| 2 |
+
Nexari Tools Engine (UPDATED)
|
| 3 |
Author: Piyush
|
| 4 |
+
Improvements:
|
| 5 |
+
- Canonical intent labels returned
|
| 6 |
+
- More robust greeting detection
|
| 7 |
+
- Safer web search formatting
|
| 8 |
+
- Defensive error handling
|
| 9 |
"""
|
| 10 |
|
| 11 |
from duckduckgo_search import DDGS
|
| 12 |
from transformers import pipeline
|
| 13 |
+
import re
|
| 14 |
|
| 15 |
print(">>> Tools: Loading Intent Classification Model...")
|
| 16 |
+
# zero-shot pipeline; if heavy, Docker predownload already handles it
|
| 17 |
intent_classifier = pipeline("zero-shot-classification", model="typeform/distilbert-base-uncased-mnli")
|
| 18 |
|
| 19 |
def analyze_intent(user_text):
|
| 20 |
"""
|
| 21 |
+
Returns canonical intent labels:
|
| 22 |
+
- "internet_search"
|
| 23 |
+
- "coding_request"
|
| 24 |
+
- "checking_time"
|
| 25 |
+
- "general"
|
| 26 |
"""
|
| 27 |
+
if not user_text:
|
| 28 |
+
return "general"
|
| 29 |
+
|
| 30 |
text_lower = user_text.lower().strip()
|
| 31 |
+
|
| 32 |
+
# === HARDCODED SAFETY (Greeting/Identity) ===
|
| 33 |
+
direct_chat_triggers = [
|
| 34 |
+
"hi", "hello", "hey", "hlo", "hola", "namaste",
|
| 35 |
+
"what is your name", "who are you", "who r you", "your name"
|
| 36 |
+
]
|
| 37 |
+
# exact match or common short greeting at start
|
| 38 |
if text_lower in direct_chat_triggers or any(text_lower.startswith(t + " ") for t in direct_chat_triggers):
|
| 39 |
+
print(f">>> Brain Override: Detected Greeting/Identity -> Force 'general'")
|
| 40 |
+
return "general"
|
| 41 |
|
| 42 |
+
# === ZERO-SHOT PASSES ===
|
| 43 |
candidate_labels = [
|
| 44 |
+
"internet search",
|
| 45 |
+
"general conversation",
|
| 46 |
+
"coding request",
|
| 47 |
"checking time"
|
| 48 |
]
|
| 49 |
+
|
| 50 |
try:
|
| 51 |
result = intent_classifier(user_text, candidate_labels)
|
| 52 |
+
top_label = result['labels'][0]
|
| 53 |
confidence = result['scores'][0]
|
| 54 |
+
print(f">>> Brain: Detected '{top_label}' ({confidence:.2f})")
|
| 55 |
+
|
|
|
|
|
|
|
| 56 |
if confidence > 0.5:
|
| 57 |
+
# map to canonical labels
|
| 58 |
+
mapping = {
|
| 59 |
+
"internet search": "internet_search",
|
| 60 |
+
"general conversation": "general",
|
| 61 |
+
"coding request": "coding_request",
|
| 62 |
+
"checking time": "checking_time"
|
| 63 |
+
}
|
| 64 |
+
return mapping.get(top_label, "general")
|
| 65 |
except Exception as e:
|
| 66 |
print(f"Intent Error: {e}")
|
|
|
|
|
|
|
| 67 |
|
| 68 |
+
return "general"
|
| 69 |
+
|
| 70 |
+
def perform_web_search(user_text, max_results=3):
|
| 71 |
"""
|
| 72 |
+
Execute a brief web search and return a compact summary.
|
| 73 |
"""
|
| 74 |
try:
|
| 75 |
clean_query = user_text.lower()
|
| 76 |
+
remove_phrases = ["search for", "google", "find", "tell me about", "latest info on", "news about", "lookup"]
|
| 77 |
for phrase in remove_phrases:
|
| 78 |
clean_query = clean_query.replace(phrase, "")
|
| 79 |
+
|
| 80 |
clean_query = clean_query.strip()
|
| 81 |
+
if len(clean_query) < 2:
|
| 82 |
+
clean_query = user_text
|
| 83 |
|
| 84 |
print(f">>> Action: Searching Web for '{clean_query}'...")
|
| 85 |
+
results = list(DDGS().text(clean_query, max_results=max_results))
|
| 86 |
+
|
| 87 |
if results:
|
| 88 |
+
parts = []
|
| 89 |
+
for r in results:
|
| 90 |
+
title = r.get('title', '').strip()
|
| 91 |
+
body = re.sub(r'\s+', ' ', r.get('body', '').strip())
|
| 92 |
+
# keep short
|
| 93 |
+
if len(body) > 250:
|
| 94 |
+
body = body[:250].rsplit(' ', 1)[0] + "..."
|
| 95 |
+
parts.append(f"- {title}: {body}")
|
| 96 |
+
summary = "\n".join(parts)
|
| 97 |
+
return f"### WEB DATA (short) ###\n{summary}\nINSTRUCTION: Use this to inform the answer; prefer concise synthesis."
|
| 98 |
return ""
|
|
|
|
| 99 |
except Exception as e:
|
| 100 |
print(f"Search Error: {e}")
|
| 101 |
+
return ""
|