Add application file

app.py

@@ -0,0 +1,120 @@
+import gradio as gr
+import pandas as pd
+import nltk
+from nltk.corpus import stopwords
+from sentence_transformers import SentenceTransformer, util
+from sklearn.feature_extraction.text import TfidfVectorizer
+from sklearn.metrics.pairwise import cosine_similarity
+
+from unsloth import FastLanguageModel
+from peft import PeftModel
+from unsloth.chat_templates import get_chat_template
+
+# Download NLTK stopwords if not already downloaded
+nltk.download("stopwords")
+
+# 1. Load model + tokenizer
+base_model_name = "unsloth/Llama-3.2-3B-Instruct"
+model, tokenizer = FastLanguageModel.from_pretrained(
+    model_name=base_model_name,
+    max_seq_length=2048,
+    dtype=None,
+    load_in_4bit=True
+)
+
+# 2. Load the LoRA adapter
+adapter_path = "jaspersands/model"  # Adjust if needed
+model = PeftModel.from_pretrained(model, adapter_path)
+
+# 3. Load data
+file_path = "Clean Missouri Data.csv"  # Ensure this CSV is in your repo
+df = pd.read_csv(file_path, encoding="MacRoman")
+
+# 4. Define helper functions
+def search_relevant_policies(query, df, top_n=10):
+    tfidf = TfidfVectorizer(stop_words='english')
+    tfidf_matrix = tfidf.fit_transform(df['Content'])
+    query_vector = tfidf.transform([query])
+    cosine_sim = cosine_similarity(query_vector, tfidf_matrix).flatten()
+    top_indices = cosine_sim.argsort()[-top_n:][::-1]
+    return df.iloc[top_indices]
+
+def get_content_after_query(response_text, query):
+    query_position = response_text.lower().find(query.lower())
+    if query_position != -1:
+        res = response_text[query_position + len(query):].strip()
+        return res[11:]
+    else:
+        return response_text.strip()
+
+def process_query(query, tokenizer):
+    # 1. Get relevant policies
+    relevant_policies = search_relevant_policies(query, df)
+
+    # 2. Format relevant policies
+    formatted_policies = []
+    for index, row in relevant_policies.iterrows():
+        formatted_policy = (
+            f"Title: {row['Title']}\nTerritory: {row['Territory']}\n"
+            f"Type: {row['Type']}\nYear: {row['Year']}\nCategory: {row['Category']}\n"
+            f"From: {row['From']}\nTo: {row['To']}\nContent: {row['Content']}\n"
+            f"Link: {row['Link to Content']}\n"
+        )
+        formatted_policies.append(formatted_policy)
+    relevant_policy_text = "\n\n".join(formatted_policies)
+
+    # 3. Create messages for model
+    messages_with_relevant_policies = [
+        {"role": "system", "content": relevant_policy_text},
+        {"role": "user", "content": query},
+    ]
+
+    # 4. Tokenize with chat template
+    tokenizer = get_chat_template(tokenizer, chat_template="llama-3.1")
+    inputs = tokenizer.apply_chat_template(
+        messages_with_relevant_policies,
+        tokenize=True,
+        add_generation_prompt=True,
+        return_tensors="pt"
+    ).to("cuda")
+
+    # 5. Generate output
+    FastLanguageModel.for_inference(model)
+    outputs = model.generate(
+        input_ids=inputs,
+        max_new_tokens=256,
+        use_cache=True,
+        temperature=1.5,
+        min_p=0.1
+    )
+    generated_response = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
+    response = get_content_after_query(generated_response, query)
+
+    # 6. Rank the top 10 policies using SBERT
+    model_sbert = SentenceTransformer("all-MiniLM-L6-v2")
+    response_embedding = model_sbert.encode(generated_response, convert_to_tensor=True)
+    policy_embeddings = model_sbert.encode(relevant_policies['Content'].tolist(), convert_to_tensor=True)
+    cosine_similarities = util.cos_sim(response_embedding, policy_embeddings).flatten()
+    most_relevant_index = cosine_similarities.argmax().item()
+    most_relevant_link = relevant_policies.iloc[most_relevant_index]['Link to Content']
+
+    return {"response": response, "most_relevant_link": most_relevant_link}
+
+# 5. Gradio interface
+def answer_query(u_query):
+    result = process_query(u_query, tokenizer)
+    return result["response"], result["most_relevant_link"]
+
+demo = gr.Interface(
+    fn=answer_query,
+    inputs="text",
+    outputs=[
+        gr.Textbox(label="System Response"),
+        gr.Textbox(label="Relevant Link")
+    ],
+    title="Foster Questions",
+    description="Enter your question about the US foster system"
+)
+
+if __name__ == "__main__":
+    demo.launch()

requirements.txt

@@ -0,0 +1,8 @@
+# requirements.txt
+unsloth
+peft
+gradio
+scikit-learn
+pandas
+nltk
+sentence-transformers