Add initial implementation of GenAI and related components for mental health assistance

.env

@@ -0,0 +1 @@
+GOOGLE_API_KEY = 'AIzaSyCuB81K0_G9wYnUkX9QF20bFSD7fEnTJ6k'

app.py

@@ -0,0 +1,93 @@
+import copy
+
+import gradio as gr
+import pandas as pd
+from datasets import load_dataset
+
+from src.genai import GenAI
+
+# from src.semantic_searcher import SemanticSearcher
+# from src.upvote_predictor import UpvotePredictor
+
+# Load the dataset
+dataset_counsel_chat = load_dataset("nbertagnolli/counsel-chat")
+df_counsel_chat = pd.DataFrame(dataset_counsel_chat["train"])
+df_counsel_chat_topic = copy.deepcopy(
+    df_counsel_chat[
+        ["questionID", "questionTitle", "questionText", "answerText", "topic"]
+    ]
+)
+df_counsel_chat_topic["questionCombined"] = df_counsel_chat_topic.apply(
+    lambda x: (
+        f"QUESTION_TITLE: {x['questionTitle']}\nQUESTION_CONTEXT: {x['questionText']}"
+    ),
+    axis=1,
+)
+df_counsel_chat_topic = df_counsel_chat_topic.drop_duplicates(
+    subset="questionID"
+).reset_index(drop=True)
+# list of unique topics
+unique_topics = sorted(df_counsel_chat_topic["topic"].unique().tolist())
+unique_topics = "\n".join(
+    [f"{idx+1}. {topic}" for idx, topic in enumerate(unique_topics)]
+)
+
+# few examples
+few_examples = (
+    df_counsel_chat_topic.groupby("topic", as_index=False)[
+        ["questionID", "questionCombined", "answerText", "topic"]
+    ]
+    .apply(lambda s: s.sample(1))
+    .reset_index(drop=True)
+)
+few_examples["examples"] = few_examples.apply(
+    lambda x: (
+        f"{x['questionCombined']}\nTOPIC: {x['topic']}\nANSWER: {x['answerText']}"
+    ),
+    axis=1,
+)
+examples = "\n".join(
+    f"<EXAMPLE {idx+1} start>\n{example}\n<EXAMPLE {idx+1} end>\n\n"
+    for idx, example in enumerate(few_examples["examples"].to_list())
+)
+
+# Initialize the SemanticSearcher
+genai = GenAI()
+# upvote_predictor = UpvotePredictor("models/bert_model")
+# _ = SemanticSearcher(df_counsel_chat_topic)
+
+
+def get_output(question: str, question_context: str = None) -> str:
+    answer, topic = genai.generate_content(
+        question, question_context, unique_topics, examples
+    )
+    return (answer, topic, "Yes", pd.DataFrame())
+    # upvote_prediction = upvote_predictor.get_upvote_prediction(
+    #     question, answer, question_context
+    # )
+    # return (answer, topic, upvote_prediction[0], upvote_prediction[1])
+
+
+demo = gr.Interface(
+    fn=get_output,
+    inputs=[
+        gr.Textbox(label="Input Question"),
+        gr.Textbox(label="(Optional) Additional Context for Question"),
+    ],
+    outputs=[
+        gr.Textbox(label="GenAI based suggestion"),
+        gr.Textbox(label="Suggested Topic of Question"),
+        gr.Textbox(label="Is GenAI based suggestion credible?"),
+        gr.Dataframe(
+            label=(
+                "Semantically similar questions (and other metadata) to input question."
+                " Will be available if GenAI based suggestion is not credible."
+            )
+        ),
+    ],
+)
+
+demo.launch(debug=True)
+# #input question
+# input_question_context = "I'm going through some things with my feelings and myself. I barely sleep and I do nothing but think about how I'm worthless and how I shouldn't be here. I've never tried or contemplated suicide. I've always wanted to fix my issues, but I never get around to it. How can I change my feeling of being worthless to everyone?"
+# input_question = "How can I change my feeling of being worthless to everyone?"

requirements.txt

@@ -0,0 +1,101 @@
+-i https://pypi.org/simple
+aiofiles==23.2.1; python_version >= '3.7'
+aiohappyeyeballs==2.6.1; python_version >= '3.9'
+aiohttp==3.11.13; python_version >= '3.9'
+aiosignal==1.3.2; python_version >= '3.9'
+annotated-types==0.7.0; python_version >= '3.8'
+anyio==4.8.0; python_version >= '3.9'
+attrs==25.3.0; python_version >= '3.8'
+cachetools==5.5.2; python_version >= '3.7'
+certifi==2025.1.31; python_version >= '3.6'
+charset-normalizer==3.4.1; python_version >= '3.7'
+click==8.1.8; python_version >= '3.7'
+datasets==3.4.0; python_full_version >= '3.9.0'
+dill==0.3.8; python_version >= '3.8'
+distro==1.9.0; python_version >= '3.6'
+fastapi==0.115.11; python_version >= '3.8'
+ffmpy==0.5.0; python_version >= '3.8' and python_version < '4.0'
+filelock==3.18.0; python_version >= '3.9'
+frozenlist==1.5.0; python_version >= '3.8'
+fsspec[http]==2024.12.0; python_version >= '3.8'
+google-ai-generativelanguage==0.6.15; python_version >= '3.7'
+google-api-core[grpc]==2.24.2; python_version >= '3.7'
+google-api-python-client==2.164.0; python_version >= '3.7'
+google-auth==2.38.0; python_version >= '3.7'
+google-auth-httplib2==0.2.0
+google-generativeai==0.8.4; python_version >= '3.9'
+googleapis-common-protos==1.69.1; python_version >= '3.7'
+gradio==5.21.0; python_version >= '3.10'
+gradio-client==1.7.2; python_version >= '3.10'
+groovy==0.1.2; python_version >= '3.10'
+grpcio==1.71.0
+grpcio-status==1.71.0
+h11==0.14.0; python_version >= '3.7'
+httpcore==1.0.7; python_version >= '3.8'
+httplib2==0.22.0; python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3'
+httpx==0.28.1; python_version >= '3.8'
+huggingface-hub==0.29.3; python_full_version >= '3.8.0'
+idna==3.10; python_version >= '3.6'
+jinja2==3.1.6; python_version >= '3.7'
+jiter==0.9.0; python_version >= '3.8'
+joblib==1.4.2; python_version >= '3.8'
+markdown-it-py==3.0.0; python_version >= '3.8'
+markupsafe==2.1.5; python_version >= '3.7'
+mdurl==0.1.2; python_version >= '3.7'
+mpmath==1.3.0
+multidict==6.1.0; python_version >= '3.8'
+multiprocess==0.70.16; python_version >= '3.8'
+networkx==3.4.2; python_version >= '3.10'
+numpy==2.2.3; python_version >= '3.10'
+openai==1.66.3; python_version >= '3.8'
+orjson==3.10.15; python_version >= '3.8'
+packaging==24.2; python_version >= '3.8'
+pandas==2.2.3; python_version >= '3.9'
+pillow==11.1.0; python_version >= '3.9'
+propcache==0.3.0; python_version >= '3.9'
+proto-plus==1.26.1; python_version >= '3.7'
+protobuf==5.29.3; python_version >= '3.8'
+pyarrow==19.0.1; python_version >= '3.9'
+pyasn1==0.6.1; python_version >= '3.8'
+pyasn1-modules==0.4.1; python_version >= '3.8'
+pydantic==2.10.6; python_version >= '3.8'
+pydantic-core==2.27.2; python_version >= '3.8'
+pydub==0.25.1
+pygments==2.19.1; python_version >= '3.8'
+pyparsing==3.2.1; python_version > '3.0'
+python-dateutil==2.9.0.post0; python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2'
+python-dotenv==1.0.1; python_version >= '3.8'
+python-multipart==0.0.20; python_version >= '3.8'
+pytz==2025.1
+pyyaml==6.0.2; python_version >= '3.8'
+regex==2024.11.6; python_version >= '3.8'
+requests==2.32.3; python_version >= '3.8'
+rich==13.9.4; python_full_version >= '3.8.0'
+rsa==4.9; python_version >= '3.6' and python_version < '4'
+ruff==0.11.0; sys_platform != 'emscripten'
+safehttpx==0.1.6; python_version >= '3.10'
+safetensors==0.5.3; python_version >= '3.7'
+scikit-learn==1.6.1; python_version >= '3.9'
+scipy==1.15.2; python_version >= '3.10'
+semantic-version==2.10.0; python_version >= '2.7'
+sentence-transformers==3.4.1; python_version >= '3.9'
+shellingham==1.5.4; python_version >= '3.7'
+six==1.17.0; python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2'
+sniffio==1.3.1; python_version >= '3.7'
+starlette==0.46.1; sys_platform != 'emscripten'
+sympy==1.13.3; python_version >= '3.8'
+threadpoolctl==3.6.0; python_version >= '3.9'
+tokenizers==0.21.1; python_version >= '3.9'
+tomlkit==0.13.2; python_version >= '3.8'
+torch==2.2.2; python_full_version >= '3.8.0'
+tqdm==4.67.1; python_version >= '3.7'
+transformers==4.49.0; python_full_version >= '3.9.0'
+typer==0.15.2; sys_platform != 'emscripten'
+typing-extensions==4.12.2; python_version >= '3.8'
+tzdata==2025.1; python_version >= '2'
+uritemplate==4.1.1; python_version >= '3.6'
+urllib3==2.3.0; python_version >= '3.9'
+uvicorn==0.34.0; sys_platform != 'emscripten'
+websockets==15.0.1; python_version >= '3.9'
+xxhash==3.5.0; python_version >= '3.7'
+yarl==1.18.3; python_version >= '3.9'

src/genai.py

@@ -0,0 +1,69 @@
+import json
+import os
+
+import google.generativeai as genai
+from dotenv import load_dotenv
+from pydantic import BaseModel
+
+load_dotenv()
+
+GOOGLE_API_KEY = os.environ.get("GOOGLE_API_KEY")
+genai.configure(api_key=GOOGLE_API_KEY)
+
+
+class OutputSchema(BaseModel):
+    topic: str
+    answer: str
+
+
+def valid_output(output: dict) -> dict:
+    return OutputSchema(**output)
+
+
+class GenAI:
+    def __init__(self):
+        self.genai_model = genai.GenerativeModel("gemini-2.0-flash-lite")
+
+    def generate_content(
+        self,
+        question: str,
+        question_context: str,
+        unique_topics: str,
+        few_shot_examples: str,
+    ):
+        prompt = f"""
+    INSTRUCTIONS:
+        1. You are a expert assistant to mental health counselors.
+        2. You are given following:
+            2.1 Input Question: A input question from mental health counselor seeking assistance.
+            2.2 Input Question Context: Additional context for input question. Can be None. If available, utilize this context in your response.
+            2.3 Topics: Categories of topics to which any input question may belong. Any input question will be categorized to one of these topics.
+            2.4 Few-shot examples: Some examples
+        3. Output the following:
+            3.1 A topic from list of topics for input question.
+            3.2 A precise answer for the input question.
+            - Length of answer should not exceed 256 words.
+        4. Your output MUST be a VALID JSON format.
+            4.1 Follow the sample JSON format given below.
+
+    INPUT:
+        Input Question: {question}
+        Input Question Context: {question_context}
+
+        Topics: {unique_topics}
+
+        Few-shot examples: {few_shot_examples}
+
+    Sample output JSON format: [{{
+        "topic": "A topic from list of topics for input question."
+        "answer": "An answer for the input question."
+        }}]
+
+    OUTPUT:"""
+        response = self.genai_model.generate_content(prompt)
+        out = response.text
+        out = out[7:-3]
+        out = json.loads(out)
+        valid_response = valid_output(out[0])
+
+        return (valid_response.answer, valid_response.topic)

src/semantic_searcher.py

@@ -0,0 +1,41 @@
+import torch
+from sentence_transformers import SentenceTransformer
+
+
+class SemanticSearcher:
+    def __init__(self, df_counsel_chat_topic):
+        self.df_counsel_chat_topic = df_counsel_chat_topic
+        self.embedder = SentenceTransformer("all-MiniLM-L6-v2")
+        self.question_embeddings = self.embedder.encode(
+            self.df_counsel_chat_topic["questionCombined"].tolist(),
+            show_progress_bar=True,
+            convert_to_tensor=True,
+        )
+
+    def retrieve_relevant_qna(self, question: str, question_context: str = None) -> str:
+        if question_context is None:
+            question_context = ""
+        query = question + "\n" + question_context
+        query_embedding = self.embedder.encode(query, convert_to_tensor=True)
+
+        # We use cosine-similarity and torch.topk to find the highest 5 scores
+        similarity_scores = self.embedder.similarity(
+            query_embedding, self.question_embeddings
+        )[0]
+        _, indices = torch.topk(similarity_scores, k=1)
+        index = indices.tolist()
+        question_id = self.df_counsel_chat_topic.loc[index, "questionID"].values[0]
+        relevant_qna = (
+            self.df_counsel_chat.loc[self.df_counsel_chat["questionID"] == question_id]
+            .sort_values(by=["upvotes", "views"], ascending=False)
+            .head(3)[[
+                "questionTitle",
+                "topic",
+                "therapistInfo",
+                "therapistURL",
+                "answerText",
+                "upvotes",
+                "views",
+            ]]
+        )
+        return relevant_qna

src/upvote_predictor.py

@@ -0,0 +1,71 @@
+import numpy as np
+import pandas as pd
+import torch
+from torch.utils.data import DataLoader, SequentialSampler, TensorDataset
+from transformers import BertTokenizer
+
+from src.semantic_searcher import retrieve_relevant_qna
+
+
+class UpvotePredictor:
+    def __init__(self, model_path: str):
+        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+        self.upvote_ml_model = torch.load(
+            model_path, map_location=torch.device("cpu"), weights_only=False
+        )
+        self.tokenizer = BertTokenizer.from_pretrained(
+            "bert-base-uncased", do_lower_case=True
+        )
+        self.upvote_ml_model.to(self.device)
+        self.upvote_ml_model.eval()
+
+    def get_upvote_prediction(
+        self, question: str, answer: str, question_context: str = None
+    ) -> int:
+        llm_response_input_ids = []
+        llm_response_attention_masks = []
+
+        encoded_dict = self.tokenizer.encode_plus(
+            answer,
+            add_special_tokens=True,
+            max_length=256,
+            padding="max_length",
+            truncation=True,
+            return_attention_mask=True,
+            return_tensors="pt",
+        )
+        llm_response_input_ids.append(encoded_dict["input_ids"])
+        llm_response_attention_masks.append(encoded_dict["attention_mask"])
+        llm_response_input_ids = torch.cat(llm_response_input_ids, dim=0)
+        llm_response_attention_masks = torch.cat(llm_response_attention_masks, dim=0)
+
+        test_dataset = TensorDataset(
+            llm_response_input_ids, llm_response_attention_masks
+        )
+        test_dataloader = DataLoader(
+            test_dataset,  # The validation samples.
+            sampler=SequentialSampler(test_dataset),  # Pull out batches sequentially.
+            batch_size=1,  # Evaluate with this batch size.
+        )
+
+        predictions = []
+        for batch in test_dataloader:
+            b_input_ids = batch[0].to(self.device)
+            b_input_mask = batch[1].to(self.device)
+            with torch.no_grad():
+                output = self.upvote_ml_model(
+                    b_input_ids, token_type_ids=None, attention_mask=b_input_mask
+                )
+                logits = output.logits
+                logits = logits.detach().cpu().numpy()
+                pred_flat = np.argmax(logits, axis=1).flatten()
+
+                predictions.extend(list(pred_flat))
+
+        if predictions[0] == 0:
+            return (
+                "Not credible suggestion",
+                retrieve_relevant_qna(question, question_context),
+            )
+        else:
+            return ("Credible suggestion", pd.DataFrame())