[#143] Double check non english detection with nltk english stop_word…

…s n words

assistant/assistant.py

@@ -3,17 +3,33 @@
 import asyncio
 import logging
 import chromadb
+import nltk
 
 from time import sleep
 from openai import OpenAI
-from langdetect import detect
+
+from langdetect import detect, DetectorFactory
+from langdetect.lang_detect_exception import LangDetectException
 from datetime import datetime, timezone
 from Akvo_rabbitmq_client import rabbitmq_client
 from typing import Optional
 from db import connect_to_sqlite, get_stable_prompt
+from nltk.corpus import words, stopwords
+
 
 logger = logging.getLogger(__name__)
 
+# Ensure reproducibility
+DetectorFactory.seed = 0
+
+# Ensure the NLTK resources are downloaded
+nltk.download("punkt_tab")
+nltk.download("words")
+nltk.download("stopwords")
+
+# Get English stopwords and the set of English words
+english_stopwords = set(stopwords.words("english"))
+english_words = set(w.lower() for w in words.words())
 
 openai = OpenAI()
 
@@ -251,7 +267,29 @@ async def publish_reliably(queue_message: str) -> None:
 
 
 def get_language(user_prompt) -> str:
-    detected_language = detect(user_prompt)
+    try:
+        # Use langdetect to get an initial guess
+        detected_language = detect(user_prompt)
+    except LangDetectException:
+        return "en"
+
+    # Double check to make sure not en
+    if detected_language != "en":
+        # Tokenize the text to get individual words
+        tokens = nltk.word_tokenize(user_prompt.lower())
+
+        # Count the number of English stopwords and valid English words
+        stopwords_count = sum(word in english_stopwords for word in tokens)
+        valid_word_count = sum(word in english_words for word in tokens)
+
+        # Heuristic: check English stopwords or words, reconsider
+        if (
+            stopwords_count > len(tokens) / 2
+            or valid_word_count > len(tokens) / 2
+        ):
+            detected_language = "en"
+
+    # check if detected language supported in env settings
     if detected_language not in assistant_data:
         logger.warning(
             f"[ASSISTANT] -> Unsupported lang detected: {detected_language}"
@@ -283,6 +321,12 @@ async def on_message(body: str) -> None:
     user_prompt = from_client["body"]
 
     detected_language = get_language(user_prompt)
+    logger.info(f"[ASSISTANT] -> user prompt: {user_prompt}")
+    logger.info(f"[ASSISTANT] -> detected lang: {detected_language}")
+    logger.info(
+        f"[ASSISTANT] -> assistant_data: {assistant_data[detected_language]}"
+    )
+
     knowledge_base = assistant_data[detected_language]["knowledge_base"]
     system_prompt = assistant_data[detected_language]["system_prompt"]
     rag_prompt = assistant_data[detected_language]["rag_prompt"]

assistant/requirements.txt

@@ -4,5 +4,6 @@ pika==1.3.2
 black==24.4.2
 flake8==7.1.0
 langdetect==1.0.9
+nltk==3.9.1
 pytest==8.3.2
 pandas==1.5.3

assistant/tests/test_assistant.py

@@ -98,6 +98,17 @@ def test_language_support():
     assert rag_prompt == sw_prompt["rag_prompt"]
     assert ragless_prompt == sw_prompt["ragless_prompt"]
 
+    detected_language = get_language("I have question about coffee")
+    assert detected_language == "en"
+    knowledge_base = assistant_data[detected_language]["knowledge_base"]
+    system_prompt = assistant_data[detected_language]["system_prompt"]
+    rag_prompt = assistant_data[detected_language]["rag_prompt"]
+    ragless_prompt = assistant_data[detected_language]["ragless_prompt"]
+    assert knowledge_base.name == "EPPO-datasheets-en"
+    assert system_prompt == en_prompt["system_prompt"]
+    assert rag_prompt == en_prompt["rag_prompt"]
+    assert ragless_prompt == en_prompt["ragless_prompt"]
+
 
 def test_query_llm():
     with patch("assistant.OpenAI") as mock_openai: