Merge branch 'main' into prod

api/backend/dataloaders/arxiv.py

@@ -1,18 +1,17 @@
 from langchain_community.document_loaders import ArxivLoader
 from langchain_core.documents.base import Document
-from transformers import pipeline
-
-TOKEN_CLASSIFIER = "ml6team/keyphrase-extraction-distilbert-inspec"
+import modal
 # yanekyuk/bert-uncased-keyword-extractor -- this model is very fast but does not always find tokens. Perhaps it is the correct choice for general questions 
 # ml6team/keyphrase-extraction-distilbert-inspec -- use this model for science-y questions: 
 
 def supplementary_info(question_text: str, answer_text: str):
 
+    find_keywords = modal.Function.lookup("find_keyword", "find_keywords")
     # convert question_text into a query for arxiv
         # extract topic from string
-    query = find_keywords(question_text)
+    query = find_keywords.remote(question_text)
     document = find_arxiv_document(query)
-
+    
 
     # convert answer_text into a query for arxi
 
@@ -26,24 +25,6 @@ def supplementary_info(question_text: str, answer_text: str):
         Would you like to hear a summary of a different article?"
         """
 
-
-def find_keywords(text) -> str:
-    """
-    If keywords are found, find_keywords returns a string containing the top three keywords from the text.
-    If no keywords are found, find_keywords returns a string containing the word "no-tokens-found".
-    """
-    pipe = pipeline("token-classification", model=TOKEN_CLASSIFIER)
-    keywords = pipe(text)
-
-    if len(keywords) == 0:
-        return "no-tokens-found"
-    if len(keywords) > 3:
-        keywords = keywords[:3]
-    print(keywords)
-    keywords = [k.get("word") for k in keywords]
-    keywords = " ".join(keywords)
-    return keywords
-
 def find_arxiv_document(query_string) -> Document:
     docs = ArxivLoader(query=query_string, load_max_docs=1).load() 
     # consider adding a layer of LLM-comparison here.
@@ -56,4 +37,4 @@ def find_arxiv_document(query_string) -> Document:
 
 
 # retrieve arxiv texts from vector db when user asks a question that would require an arxiv text
-# ^^ do the same thing for topics that require data from other sources
+# ^^ do the same thing for topics that require data from other sources

api/backend/handlers/answer_question.py

@@ -23,17 +23,18 @@ def answer_question_handler(question_audio: IO, data: QuestionTextRequestDto):
     ask_question = build_ask_question()
     answer_text = ask_question(question_text, data.text)
 
-    # question_type = classify_question_type(question_text)
-    # supplementary_info = build_supplementary_info(question_type)
-    # followup_answer = supplementary_info(question_text, answer_text)
+    question_type = classify_question_type(question_text)
+    supplementary_info = build_supplementary_info(question_type)
+    followup_answer = supplementary_info(question_text, answer_text)
+    print(followup_answer)
     # add follow-up answer:
         # "here is an arxiv article related to your question.
         # The title is "". The summary is "" and the authors are "".
         # Would you like to hear the article?
         # Or would you like to hear a summary of a different article"
 
     text_to_speech = build_text_to_speech()
-    audio_content = text_to_speech(answer_text, data.emotion, data.speed)
+    audio_content = text_to_speech(answer_text + followup_answer, data.emotion, data.speed)
 
     file_contents = io.BytesIO(audio_content)
     file_contents.seek(0)

api/backend/handlers/find_keywords.py

@@ -0,0 +1,56 @@
+import modal
+
+app = modal.App(
+    "find_keyword"
+) 
+
+def download_model_to_image(model_dir, model_name):
+    import os
+    from huggingface_hub import snapshot_download
+    from transformers.utils import move_cache
+
+    os.makedirs(model_dir, exist_ok=True)
+
+    snapshot_download(
+        model_name,
+        local_dir=model_dir
+    )
+    move_cache()
+
+TOKEN_CLASSIFIER = "ml6team/keyphrase-extraction-distilbert-inspec"
+MODEL_DIR = "/token_classifier"
+transformers_image = (
+    modal.Image.debian_slim()
+    .pip_install(
+        "torch==2.2.2",
+        "transformers==4.40.0",
+        "hf-transfer==0.1.6",
+        "huggingface_hub==0.22.2",
+    )
+    .run_function(download_model_to_image,
+        timeout=60 * 4,
+        kwargs={
+            "model_dir": MODEL_DIR,
+            "model_name": TOKEN_CLASSIFIER,
+        }
+    )
+)
+
+@app.function(image=transformers_image, gpu="a10g")
+def find_keywords(text) -> str:
+    """
+    If keywords are found, find_keywords returns a string containing the top three keywords from the text.
+    If no keywords are found, find_keywords returns a string containing the word "no-tokens-found".
+    """
+    from transformers import pipeline
+    pipe = pipeline("token-classification", model=MODEL_DIR)
+    keywords = pipe(text)
+
+    if len(keywords) == 0:
+        return "no-tokens-found"
+    if len(keywords) > 3:
+        keywords = keywords[:3]
+    print(keywords)
+    keywords = [k.get("word") for k in keywords]
+    keywords = " ".join(keywords)
+    return keywords

api/cloud/Dockerfile

@@ -24,7 +24,7 @@ RUN groupadd --gid 1010 socketWriters
 # create appUser
 # gid 1010 is socketWriters
 # uid 1007 will own the /app dir
-# Coqui saves models to the user's home dir, so we need to create a user with a home dir
+# we need to create a user with a home dir
     && useradd -u 1007 -g 1010 -m -p "$(cat password)" appUser                  \ 
     && shred -u password                                                        \
 # set HOME on non-login user changes for appUser:

api/cloud/nginx.conf

@@ -28,6 +28,7 @@ http {
       ''      close;
     }
 
+    # keep this in place so that we can dev against the docker container locally
     server {
         listen 80;
         server_name drive-gooder.com;
@@ -36,7 +37,7 @@ http {
 
     server {
         listen 3000;
-        listen 443 ssl;
+        listen 443 ssl; # keep this line so that we can dev against the docker container locally
         server_name drive-gooder.com;
 
         proxy_ssl_server_name on;