Init app

app.py

@@ -1,9 +1,8 @@
-import torch
 from fastapi import FastAPI, File, UploadFile, Form
-from transformers import GPT2Tokenizer, GPT2Model
-from sklearn.metrics.pairwise import cosine_similarity
+from llama_cpp import Llama
 from typing import List
 from fastapi.middleware.cors import CORSMiddleware
+import json
 
 # Initialize FastAPI app
 app = FastAPI()
@@ -17,21 +16,11 @@ app.add_middleware(
     allow_headers=["*"],
 )
 
-# Load GPT-2 model and tokenizer
-tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
-model = GPT2Model.from_pretrained("gpt2")
-
-# Helper function to get GPT-2 embeddings
-def get_gpt2_embedding(text):
-    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)
-    outputs = model(**inputs)
-    embeddings = torch.mean(outputs.last_hidden_state, dim=1).detach().numpy()
-    return embeddings
-
-# Helper function to calculate cosine similarity
-def calculate_similarity(embedding1, embedding2):
-    similarity = cosine_similarity(embedding1, embedding2)
-    return similarity[0][0]
+# Load the Llama model
+llm = Llama.from_pretrained(
+    repo_id="HuggingFaceTB/SmolLM2-1.7B-Instruct-GGUF",
+    filename="smollm2-1.7b-instruct-q4_k_m.gguf",
+)
 
 # Endpoint to upload CV file and store CV text
 @app.post("/upload-cv/")
@@ -43,22 +32,29 @@ async def upload_cv(file: UploadFile = File(...)):
 # Endpoint to compare job descriptions with the CV text
 @app.post("/compare/")
 async def compare_job_cv(job_descriptions: str = Form(...), cv_text: str = Form(...)):
-    # Generate embedding for the CV text
-    cv_embedding = get_gpt2_embedding(cv_text)
-    
-    # Split job descriptions by line and calculate scores
+    # Split job descriptions by line
     descriptions = job_descriptions.strip().split("\n")
     results = []
     
     for description in descriptions:
-        job_embedding = get_gpt2_embedding(description)
-        similarity_score = calculate_similarity(cv_embedding, job_embedding)
+        # Create chat messages to prompt Llama for each job description
+        messages = [
+            {"role": "user", "content": f"Compare the following job description with this resume. Job Description: {description}. Resume: {cv_text}. Give a match score and brief analysis."}
+        ]
+        
+        # Generate response using Llama
+        response = llm.create_chat_completion(messages=messages)
+        response_content = response["choices"][0]["message"]["content"]
         
-        # Append results
-        results.append({
-            "Job Description": description,
-            "Match Score": round(similarity_score, 2)
-        })
+        # Parse response content for a score and summary
+        try:
+            response_data = json.loads(response_content)
+            results.append(response_data)
+        except json.JSONDecodeError:
+            results.append({
+                "Job Description": description,
+                "Analysis": response_content  # Use raw response if JSON parsing fails
+            })
     
     return {"results": results}
 

requirements.txt

@@ -3,4 +3,5 @@ uvicorn
 torch
 transformers
 scikit-learn
-peft
+peft
+git+https://github.com/abetlen/llama-cpp-python.git