Spaces:
Sleeping
Sleeping
update-fix/Fixed major issues of code, api, and models. Model were updated with proper formatiing, FastAPI code is simpler and cleaner, finally the api is optimized by removing pdf parsings
Browse files- main.py +55 -138
- models/mnr_1/README.md +2 -2
- models/mnr_1/model.safetensors +1 -1
- models/qwen-2.5-v2/checkpoint-1010/adapter_model.safetensors +0 -3
- models/{qwen-2.5-v2/checkpoint-1010 → qwen_final}/README.md +0 -0
- models/{qwen-2.5-v2/checkpoint-1010 → qwen_final}/adapter_config.json +2 -2
- models/{qwen-2.5-v2/README.md → qwen_final/adapter_model.safetensors} +2 -2
- models/{qwen-2.5-v2/checkpoint-1010 → qwen_final}/chat_template.jinja +0 -0
- models/{qwen-2.5-v2/checkpoint-1010 → qwen_final}/optimizer.pt +2 -2
- models/{qwen-2.5-v2/checkpoint-1010 → qwen_final}/rng_state.pth +1 -1
- models/{qwen-2.5-v2/checkpoint-1010 → qwen_final}/scheduler.pt +1 -1
- models/{qwen-2.5-v2/checkpoint-1010 → qwen_final}/tokenizer.json +0 -0
- models/{qwen-2.5-v2/checkpoint-1010 → qwen_final}/tokenizer_config.json +0 -0
- models/{qwen-2.5-v2/checkpoint-1010 → qwen_final}/trainer_state.json +2 -2
- models/{qwen-2.5-v2/checkpoint-1010 → qwen_final}/training_args.bin +1 -1
- modules/comparison.py +5 -87
- modules/pdf_to_json.py +0 -162
- modules/resume_parser.py +41 -0
- requirements.txt +0 -3
main.py
CHANGED
|
@@ -1,53 +1,38 @@
|
|
| 1 |
-
|
| 2 |
-
|
| 3 |
-
|
| 4 |
from pathlib import Path
|
|
|
|
|
|
|
|
|
|
| 5 |
|
| 6 |
-
|
| 7 |
-
from modules.comparison import CompareData
|
| 8 |
-
|
| 9 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 10 |
from peft import PeftModel
|
| 11 |
-
from sentence_transformers import SentenceTransformer
|
| 12 |
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
THRESHOLD = 0.675
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
# ---------------------------------------------
|
| 19 |
-
# 🧠 RESUME STEP LOGGER
|
| 20 |
-
# ---------------------------------------------
|
| 21 |
-
def log_resume_steps(filename: str, steps: list):
|
| 22 |
-
print("\n" + "=" * 60)
|
| 23 |
-
print(f"📄 RESUME: {filename}")
|
| 24 |
-
print("=" * 60)
|
| 25 |
-
|
| 26 |
-
for i, step in enumerate(steps, 1):
|
| 27 |
-
print(f"{i}. {step}")
|
| 28 |
|
| 29 |
-
|
|
|
|
|
|
|
| 30 |
|
|
|
|
| 31 |
|
| 32 |
-
# ---------------------------------------------
|
| 33 |
-
# 🚀 STARTUP: LOAD ALL MODELS ONCE
|
| 34 |
-
# ---------------------------------------------
|
| 35 |
@app.on_event("startup")
|
| 36 |
def load_models():
|
| 37 |
-
print("📦 Loading ALL models...")
|
| 38 |
-
|
| 39 |
BASE_DIR = Path(__file__).resolve().parent
|
| 40 |
|
| 41 |
# ===== QWEN =====
|
| 42 |
base_model_name = "Qwen/Qwen2.5-0.5B-Instruct"
|
| 43 |
-
adapter_path = BASE_DIR / "models/
|
| 44 |
|
| 45 |
tokenizer = AutoTokenizer.from_pretrained(base_model_name)
|
| 46 |
tokenizer.pad_token = tokenizer.eos_token
|
| 47 |
|
| 48 |
base_model = AutoModelForCausalLM.from_pretrained(
|
| 49 |
base_model_name,
|
| 50 |
-
|
| 51 |
device_map="cpu"
|
| 52 |
)
|
| 53 |
|
|
@@ -58,8 +43,8 @@ def load_models():
|
|
| 58 |
minilm_model = SentenceTransformer(str(BASE_DIR / "models/mnr_1"))
|
| 59 |
|
| 60 |
# ===== Inject =====
|
| 61 |
-
app.state.
|
| 62 |
-
app.state.
|
| 63 |
|
| 64 |
print("✅ All models loaded once")
|
| 65 |
|
|
@@ -75,112 +60,44 @@ def home():
|
|
| 75 |
# ---------------------------------------------
|
| 76 |
# MAIN ENDPOINT
|
| 77 |
# ---------------------------------------------
|
|
|
|
|
|
|
|
|
|
|
|
|
| 78 |
@app.post("/match-resumes/")
|
| 79 |
-
async def match_resumes(
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
steps.append(f"Received file: {file.filename}")
|
| 110 |
-
|
| 111 |
-
try:
|
| 112 |
-
# -------------------------
|
| 113 |
-
# SAVE FILE
|
| 114 |
-
# -------------------------
|
| 115 |
-
with open(file_path, "wb") as f:
|
| 116 |
-
content = await file.read()
|
| 117 |
-
f.write(content)
|
| 118 |
-
|
| 119 |
-
steps.append("PDF saved to temp storage")
|
| 120 |
-
|
| 121 |
-
# -------------------------
|
| 122 |
-
# EXTRACT TEXT
|
| 123 |
-
# -------------------------
|
| 124 |
-
resume_texts = pdf_to_json.extract_text([file_path])
|
| 125 |
-
steps.append(f"Text extracted (size={len(resume_texts)})")
|
| 126 |
-
|
| 127 |
-
# -------------------------
|
| 128 |
-
# JSON GENERATION
|
| 129 |
-
# -------------------------
|
| 130 |
-
resume_json = pdf_to_json.generate_json(resume_texts)[0]
|
| 131 |
-
|
| 132 |
-
if "error" in resume_json:
|
| 133 |
-
steps.append("❌ JSON generation failed — skipping resume")
|
| 134 |
-
log_resume_steps(file.filename, steps)
|
| 135 |
-
continue
|
| 136 |
-
|
| 137 |
-
steps.append("Structured JSON generated from resume")
|
| 138 |
-
|
| 139 |
-
# -------------------------
|
| 140 |
-
# PROCESS RESUME
|
| 141 |
-
# -------------------------
|
| 142 |
-
processed_resume = compare.process_resume(resume_json)
|
| 143 |
-
steps.append("Resume text processed/normalized")
|
| 144 |
-
|
| 145 |
-
# -------------------------
|
| 146 |
-
# ENCODING
|
| 147 |
-
# -------------------------
|
| 148 |
-
resume_emb = compare.encode(processed_resume)
|
| 149 |
-
steps.append("Resume embedding created")
|
| 150 |
-
|
| 151 |
-
# -------------------------
|
| 152 |
-
# SIMILARITY
|
| 153 |
-
# -------------------------
|
| 154 |
-
score = compare.compare(resume_emb, jd_emb)
|
| 155 |
-
steps.append(f"Similarity score computed: {score:.4f}")
|
| 156 |
-
|
| 157 |
-
# -------------------------
|
| 158 |
-
# DECISION
|
| 159 |
-
# -------------------------
|
| 160 |
-
if score >= THRESHOLD:
|
| 161 |
-
steps.append(f"✅ MATCHED (threshold={THRESHOLD})")
|
| 162 |
-
|
| 163 |
-
results.append({
|
| 164 |
-
"name": resume_json.get("name", "Unknown"),
|
| 165 |
-
"score": float(score),
|
| 166 |
-
"resume": resume_json
|
| 167 |
-
})
|
| 168 |
-
else:
|
| 169 |
-
steps.append(f"❌ REJECTED (below threshold {THRESHOLD})")
|
| 170 |
-
|
| 171 |
-
# FINAL LOG FOR THIS RESUME
|
| 172 |
-
log_resume_steps(file.filename, steps)
|
| 173 |
-
|
| 174 |
-
finally:
|
| 175 |
-
if os.path.exists(file_path):
|
| 176 |
-
os.remove(file_path)
|
| 177 |
-
steps.append("Temp file deleted")
|
| 178 |
-
|
| 179 |
-
# ---------------------------------
|
| 180 |
-
# RESPONSE
|
| 181 |
-
# ---------------------------------
|
| 182 |
return {
|
| 183 |
-
"
|
| 184 |
-
"
|
| 185 |
-
"
|
|
|
|
|
|
|
| 186 |
}
|
|
|
|
| 1 |
+
#import torch and fast api
|
| 2 |
+
import torch
|
| 3 |
+
from fastapi import FastAPI
|
| 4 |
from pathlib import Path
|
| 5 |
+
from pydantic import BaseModel
|
| 6 |
+
from typing import List
|
| 7 |
+
import time
|
| 8 |
|
| 9 |
+
#import transformer modules
|
|
|
|
|
|
|
| 10 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 11 |
from peft import PeftModel
|
|
|
|
| 12 |
|
| 13 |
+
#import sentence_transformers
|
| 14 |
+
from sentence_transformers import SentenceTransformer
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 15 |
|
| 16 |
+
# import modules
|
| 17 |
+
from modules.resume_parser import ResumeParser
|
| 18 |
+
from modules.comparison import Comparison
|
| 19 |
|
| 20 |
+
app = FastAPI()
|
| 21 |
|
|
|
|
|
|
|
|
|
|
| 22 |
@app.on_event("startup")
|
| 23 |
def load_models():
|
|
|
|
|
|
|
| 24 |
BASE_DIR = Path(__file__).resolve().parent
|
| 25 |
|
| 26 |
# ===== QWEN =====
|
| 27 |
base_model_name = "Qwen/Qwen2.5-0.5B-Instruct"
|
| 28 |
+
adapter_path = BASE_DIR / "models/qwen_final"
|
| 29 |
|
| 30 |
tokenizer = AutoTokenizer.from_pretrained(base_model_name)
|
| 31 |
tokenizer.pad_token = tokenizer.eos_token
|
| 32 |
|
| 33 |
base_model = AutoModelForCausalLM.from_pretrained(
|
| 34 |
base_model_name,
|
| 35 |
+
dtype=torch.float32,
|
| 36 |
device_map="cpu"
|
| 37 |
)
|
| 38 |
|
|
|
|
| 43 |
minilm_model = SentenceTransformer(str(BASE_DIR / "models/mnr_1"))
|
| 44 |
|
| 45 |
# ===== Inject =====
|
| 46 |
+
app.state.resume_parser = ResumeParser(qwen_model, tokenizer)
|
| 47 |
+
app.state.comparison = Comparison(minilm_model)
|
| 48 |
|
| 49 |
print("✅ All models loaded once")
|
| 50 |
|
|
|
|
| 60 |
# ---------------------------------------------
|
| 61 |
# MAIN ENDPOINT
|
| 62 |
# ---------------------------------------------
|
| 63 |
+
class ResumeMatchRequest(BaseModel):
|
| 64 |
+
job_description:str
|
| 65 |
+
resume: str
|
| 66 |
+
|
| 67 |
@app.post("/match-resumes/")
|
| 68 |
+
async def match_resumes(request:ResumeMatchRequest):
|
| 69 |
+
start = time.time()
|
| 70 |
+
resume = request.resume
|
| 71 |
+
job_description = request.job_description
|
| 72 |
+
|
| 73 |
+
#load parser
|
| 74 |
+
parser = app.state.resume_parser
|
| 75 |
+
|
| 76 |
+
#resumes in parsed format
|
| 77 |
+
resume = parser.format_data(resume)
|
| 78 |
+
|
| 79 |
+
#resumes after generated
|
| 80 |
+
resume = parser.parse(resume)
|
| 81 |
+
splitted = resume.split("\n")
|
| 82 |
+
name = splitted[0].split("name: ").pop()
|
| 83 |
+
parsed_resume = "\n".join(splitted[1:])
|
| 84 |
+
|
| 85 |
+
|
| 86 |
+
#load comparison
|
| 87 |
+
comparison = app.state.comparison
|
| 88 |
+
|
| 89 |
+
# use all mini lm to encode
|
| 90 |
+
with torch.no_grad():
|
| 91 |
+
jd = comparison.encode(job_description)
|
| 92 |
+
res = comparison.encode(parsed_resume)
|
| 93 |
+
score = comparison.compare(res, jd)
|
| 94 |
+
|
| 95 |
+
end = time.time()
|
| 96 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 97 |
return {
|
| 98 |
+
"scores":score,
|
| 99 |
+
"resume":{"name": name, "resume":parsed_resume},
|
| 100 |
+
"threshold":0.670,
|
| 101 |
+
"process_time": f"{end-start} seconds",
|
| 102 |
+
"message": "Candidate Accepted" if score>=0.670 else "Candidate Rejected"
|
| 103 |
}
|
models/mnr_1/README.md
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7dee2fd076b7c08ea50f40e7671b64e7d0eeef3941772ab7276d7c910e617fef
|
| 3 |
+
size 34755
|
models/mnr_1/model.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 90864192
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:324b8a60d2ac42d9df4fc7501daf734df6bb24fb05613f3fc0032f39659c6f0e
|
| 3 |
size 90864192
|
models/qwen-2.5-v2/checkpoint-1010/adapter_model.safetensors
DELETED
|
@@ -1,3 +0,0 @@
|
|
| 1 |
-
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:8ba09c8697ba3535b4c56bb91d91acd5eb572bc0cc86c8013a8044b7150ec026
|
| 3 |
-
size 17640136
|
|
|
|
|
|
|
|
|
|
|
|
models/{qwen-2.5-v2/checkpoint-1010 → qwen_final}/README.md
RENAMED
|
File without changes
|
models/{qwen-2.5-v2/checkpoint-1010 → qwen_final}/adapter_config.json
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d704aad32dec0f1d225c2b00c6be7ee72578e9c1402c7e6ab09ea7c57c694f31
|
| 3 |
+
size 1260
|
models/{qwen-2.5-v2/README.md → qwen_final/adapter_model.safetensors}
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:4eac848e319ef60c95caddca849d132d1d960a9f587d38a8f447c6b2a191d70e
|
| 3 |
+
size 8676008
|
models/{qwen-2.5-v2/checkpoint-1010 → qwen_final}/chat_template.jinja
RENAMED
|
File without changes
|
models/{qwen-2.5-v2/checkpoint-1010 → qwen_final}/optimizer.pt
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3b0fc672dfc6df371804814f346759d0fe3ac9f9fed155e9c8480070fe308bda
|
| 3 |
+
size 17465739
|
models/{qwen-2.5-v2/checkpoint-1010 → qwen_final}/rng_state.pth
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 14645
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:dee670dbbb7f9ae79334dd2a96a883c819ee2429d4cad998dcb558b1b2923cb8
|
| 3 |
size 14645
|
models/{qwen-2.5-v2/checkpoint-1010 → qwen_final}/scheduler.pt
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 1465
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3af8416f59a652f82564cd7a9c900059bc941731a17bd7ebfd118c6f498f6c5f
|
| 3 |
size 1465
|
models/{qwen-2.5-v2/checkpoint-1010 → qwen_final}/tokenizer.json
RENAMED
|
File without changes
|
models/{qwen-2.5-v2/checkpoint-1010 → qwen_final}/tokenizer_config.json
RENAMED
|
File without changes
|
models/{qwen-2.5-v2/checkpoint-1010 → qwen_final}/trainer_state.json
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:5801cb328e9f74599a1f9b59084068f82357990cda00ad58dd602ca9853415ab
|
| 3 |
+
size 19095
|
models/{qwen-2.5-v2/checkpoint-1010 → qwen_final}/training_args.bin
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 5585
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:abd69ce1cc8803e63919dc90432c3635c95ef64f5de7b931f16e916397248fea
|
| 3 |
size 5585
|
modules/comparison.py
CHANGED
|
@@ -1,93 +1,11 @@
|
|
| 1 |
-
from sentence_transformers import SentenceTransformer
|
| 2 |
-
from pathlib import Path
|
| 3 |
-
import torch
|
| 4 |
import torch.nn.functional as F
|
| 5 |
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
class CompareData:
|
| 9 |
-
"""
|
| 10 |
-
Converts resume and job description JSON into embeddings
|
| 11 |
-
and compares them using PyTorch cosine similarity.
|
| 12 |
-
"""
|
| 13 |
-
|
| 14 |
def __init__(self, model):
|
| 15 |
-
|
| 16 |
self.model = model
|
| 17 |
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
# ---------------------------------------------
|
| 21 |
-
def process_resume(self, resume_json):
|
| 22 |
-
resume_text = f"""
|
| 23 |
-
[RESUME]
|
| 24 |
-
Role: {resume_json.get("role", "")}
|
| 25 |
-
|
| 26 |
-
Key Skills: {", ".join(resume_json.get("skills", []))}
|
| 27 |
-
Skills: {", ".join(resume_json.get("skills", []))}
|
| 28 |
-
|
| 29 |
-
Experience: {resume_json.get("years_of_experience", 0)} years
|
| 30 |
-
|
| 31 |
-
Work Experience:
|
| 32 |
-
{" . ".join(resume_json.get("work_experience", []))}
|
| 33 |
-
|
| 34 |
-
Education:
|
| 35 |
-
{", ".join(resume_json.get("education", []))}
|
| 36 |
-
"""
|
| 37 |
-
return resume_text.strip()
|
| 38 |
-
|
| 39 |
-
# ---------------------------------------------
|
| 40 |
-
# JOB DESCRIPTION PROCESSING
|
| 41 |
-
# ---------------------------------------------
|
| 42 |
-
def process_jd(self, jd_json):
|
| 43 |
-
jd_text = f"""
|
| 44 |
-
[JOB]
|
| 45 |
-
Role: {jd_json.get("job_title", "")}
|
| 46 |
-
|
| 47 |
-
Key Skills: {", ".join(jd_json.get("required_skills", []))}
|
| 48 |
-
Skills: {", ".join(jd_json.get("required_skills", []))}
|
| 49 |
-
|
| 50 |
-
Requirements:
|
| 51 |
-
{" . ".join(jd_json.get("requirements", []))}
|
| 52 |
-
|
| 53 |
-
Responsibilities:
|
| 54 |
-
{" . ".join(jd_json.get("responsibilities", []))}
|
| 55 |
-
|
| 56 |
-
Education:
|
| 57 |
-
{", ".join(jd_json.get("education", []))}
|
| 58 |
-
"""
|
| 59 |
-
return jd_text.strip()
|
| 60 |
-
|
| 61 |
-
# ---------------------------------------------
|
| 62 |
-
# EMBEDDING GENERATION
|
| 63 |
-
# ---------------------------------------------
|
| 64 |
-
def encode(self, text):
|
| 65 |
-
"""
|
| 66 |
-
Convert text into embedding tensor.
|
| 67 |
-
"""
|
| 68 |
-
return (
|
| 69 |
-
self.model.encode(
|
| 70 |
-
text,
|
| 71 |
-
convert_to_tensor=True,
|
| 72 |
-
normalize_embeddings=True
|
| 73 |
-
)
|
| 74 |
-
)
|
| 75 |
-
|
| 76 |
-
# ---------------------------------------------
|
| 77 |
-
# SIMILARITY CHECK (PYTORCH)
|
| 78 |
-
# ---------------------------------------------
|
| 79 |
-
def compare(self, resume_embedding, jd_embedding):
|
| 80 |
-
"""
|
| 81 |
-
Compute cosine similarity using PyTorch.
|
| 82 |
-
"""
|
| 83 |
-
|
| 84 |
-
# Ensure correct shape: [1, dim]
|
| 85 |
-
resume_embedding = resume_embedding.unsqueeze(0) if resume_embedding.dim() == 1 else resume_embedding
|
| 86 |
-
jd_embedding = jd_embedding.unsqueeze(0) if jd_embedding.dim() == 1 else jd_embedding
|
| 87 |
-
|
| 88 |
-
similarity = F.cosine_similarity(
|
| 89 |
-
resume_embedding,
|
| 90 |
-
jd_embedding
|
| 91 |
-
)
|
| 92 |
|
| 93 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
import torch.nn.functional as F
|
| 2 |
|
| 3 |
+
class Comparison:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 4 |
def __init__(self, model):
|
|
|
|
| 5 |
self.model = model
|
| 6 |
|
| 7 |
+
def encode(self, texts):
|
| 8 |
+
return self.model.encode(texts, convert_to_tensor=True, normalize_embeddings=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 9 |
|
| 10 |
+
def compare(self, resume, jd):
|
| 11 |
+
return F.cosine_similarity(jd, resume, dim=-1).tolist()
|
modules/pdf_to_json.py
DELETED
|
@@ -1,162 +0,0 @@
|
|
| 1 |
-
import os
|
| 2 |
-
import re
|
| 3 |
-
import json
|
| 4 |
-
import torch
|
| 5 |
-
from pathlib import Path
|
| 6 |
-
from pypdf import PdfReader
|
| 7 |
-
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 8 |
-
from peft import PeftModel
|
| 9 |
-
|
| 10 |
-
# =========================================================
|
| 11 |
-
# ⚙️ CPU OPTIMIZATION (IMPORTANT for HF Spaces)
|
| 12 |
-
# =========================================================
|
| 13 |
-
torch.set_num_threads(max(1, os.cpu_count() - 1))
|
| 14 |
-
os.environ["TOKENIZERS_PARALLELISM"] = "false"
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
# =========================================================
|
| 18 |
-
# 🧼 TEXT CLEANING
|
| 19 |
-
# =========================================================
|
| 20 |
-
def clean_resume_text(text: str) -> str:
|
| 21 |
-
text = re.sub(r"[\x00-\x1F\x7F]", " ", text)
|
| 22 |
-
text = text.replace("•", "-").replace("●", "-").replace("▪", "-")
|
| 23 |
-
text = re.sub(r"[ \t]+", " ", text)
|
| 24 |
-
text = re.sub(r"\n{3,}", "\n\n", text)
|
| 25 |
-
return text.strip()
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
# =========================================================
|
| 29 |
-
# 🧠 STRUCTURE ENFORCEMENT
|
| 30 |
-
# =========================================================
|
| 31 |
-
def enforce_structure(text: str) -> str:
|
| 32 |
-
sections = ["ABOUT", "EXPERIENCE", "EDUCATION", "SKILLS", "PROJECTS", "CERTIFICATIONS"]
|
| 33 |
-
|
| 34 |
-
for s in sections:
|
| 35 |
-
text = re.sub(
|
| 36 |
-
rf"\b{s}\b",
|
| 37 |
-
f"\n\n===== {s} =====\n",
|
| 38 |
-
text,
|
| 39 |
-
flags=re.I
|
| 40 |
-
)
|
| 41 |
-
|
| 42 |
-
return text
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
# =========================================================
|
| 46 |
-
# 🧾 PROMPT BUILDER
|
| 47 |
-
# =========================================================
|
| 48 |
-
def build_prompt(input_text: str) -> str:
|
| 49 |
-
schema = {
|
| 50 |
-
"name": None,
|
| 51 |
-
"email": None,
|
| 52 |
-
"location": None,
|
| 53 |
-
"phone": None,
|
| 54 |
-
"linkedin": None,
|
| 55 |
-
"github": None,
|
| 56 |
-
"work_experience": [],
|
| 57 |
-
"skills": [],
|
| 58 |
-
"projects": [],
|
| 59 |
-
"education": [],
|
| 60 |
-
"certifications": []
|
| 61 |
-
}
|
| 62 |
-
|
| 63 |
-
return f"""<|im_start|>system
|
| 64 |
-
You are a strict JSON extraction engine for resumes.
|
| 65 |
-
|
| 66 |
-
RULES:
|
| 67 |
-
- Output ONLY valid JSON
|
| 68 |
-
- No explanation
|
| 69 |
-
- No markdown
|
| 70 |
-
- Follow schema exactly
|
| 71 |
-
- Use null if missing
|
| 72 |
-
- Use [] for missing lists
|
| 73 |
-
<|im_end|>
|
| 74 |
-
|
| 75 |
-
<|im_start|>user
|
| 76 |
-
Extract resume data strictly.
|
| 77 |
-
|
| 78 |
-
Schema:
|
| 79 |
-
{json.dumps(schema, indent=2)}
|
| 80 |
-
|
| 81 |
-
RESUME:
|
| 82 |
-
{input_text}
|
| 83 |
-
<|im_end|>
|
| 84 |
-
"""
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
# =========================================================
|
| 89 |
-
# 📄 MAIN PIPELINE CLASS
|
| 90 |
-
# =========================================================
|
| 91 |
-
class PDFToJSON:
|
| 92 |
-
|
| 93 |
-
def __init__(self, model, tokenizer):
|
| 94 |
-
self.model = model
|
| 95 |
-
self.tokenizer = tokenizer
|
| 96 |
-
|
| 97 |
-
# -----------------------------------------------------
|
| 98 |
-
# 📄 PDF → TEXT
|
| 99 |
-
# -----------------------------------------------------
|
| 100 |
-
def extract_text(self, resume_pdfs):
|
| 101 |
-
cleaned_resumes = []
|
| 102 |
-
|
| 103 |
-
for pdf in resume_pdfs:
|
| 104 |
-
reader = PdfReader(pdf)
|
| 105 |
-
text_pages = []
|
| 106 |
-
|
| 107 |
-
for page in reader.pages:
|
| 108 |
-
text = page.extract_text() or ""
|
| 109 |
-
if text.strip():
|
| 110 |
-
text_pages.append(text)
|
| 111 |
-
|
| 112 |
-
full_text = "\n".join(text_pages)
|
| 113 |
-
cleaned = clean_resume_text(full_text)
|
| 114 |
-
structured = enforce_structure(cleaned)
|
| 115 |
-
|
| 116 |
-
cleaned_resumes.append(structured)
|
| 117 |
-
|
| 118 |
-
return cleaned_resumes
|
| 119 |
-
|
| 120 |
-
# -----------------------------------------------------
|
| 121 |
-
# 🧠 TEXT → JSON
|
| 122 |
-
# -----------------------------------------------------
|
| 123 |
-
def generate_json(self, resumes):
|
| 124 |
-
results = []
|
| 125 |
-
|
| 126 |
-
for resume in resumes:
|
| 127 |
-
|
| 128 |
-
prompt = build_prompt(resume)
|
| 129 |
-
|
| 130 |
-
inputs = self.tokenizer(
|
| 131 |
-
prompt,
|
| 132 |
-
return_tensors="pt",
|
| 133 |
-
truncation=True,
|
| 134 |
-
max_length=1800,
|
| 135 |
-
padding=True
|
| 136 |
-
)
|
| 137 |
-
|
| 138 |
-
with torch.no_grad():
|
| 139 |
-
output = self.model.generate(
|
| 140 |
-
**inputs,
|
| 141 |
-
max_new_tokens=500,
|
| 142 |
-
do_sample=False,
|
| 143 |
-
eos_token_id=self.tokenizer.eos_token_id,
|
| 144 |
-
repetition_penalty=1.05,
|
| 145 |
-
)
|
| 146 |
-
|
| 147 |
-
input_len = inputs["input_ids"].shape[1]
|
| 148 |
-
output = output[0][input_len:]
|
| 149 |
-
|
| 150 |
-
response = self.tokenizer.decode(output, skip_special_tokens=True)
|
| 151 |
-
|
| 152 |
-
try:
|
| 153 |
-
match = re.search(r"\{[\s\S]*\}", response)
|
| 154 |
-
parsed = json.loads(match.group())
|
| 155 |
-
results.append(parsed)
|
| 156 |
-
except:
|
| 157 |
-
results.append({
|
| 158 |
-
"error": "parse error",
|
| 159 |
-
"raw_output": response[:300]
|
| 160 |
-
})
|
| 161 |
-
|
| 162 |
-
return results
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
modules/resume_parser.py
ADDED
|
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
|
| 3 |
+
class ResumeParser:
|
| 4 |
+
def __init__(self, model, tokenizer):
|
| 5 |
+
self.model = model
|
| 6 |
+
self.tokenizer = tokenizer
|
| 7 |
+
|
| 8 |
+
def format_data(self, txt):
|
| 9 |
+
prompt = f"""<|im_start|>system
|
| 10 |
+
Extract and format resume: name | role | years | skills | core skills | experiences | education | certifications.
|
| 11 |
+
Extract ONLY from input. NO inference or additions.
|
| 12 |
+
<|im_end|>
|
| 13 |
+
|
| 14 |
+
<|im_start|>user
|
| 15 |
+
{txt}
|
| 16 |
+
<|im_end|>
|
| 17 |
+
|
| 18 |
+
<|im_start|>assistant
|
| 19 |
+
"""
|
| 20 |
+
return prompt
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
def parse(self, resume):
|
| 24 |
+
|
| 25 |
+
inputs = self.tokenizer(resume, return_tensors="pt", padding=False, truncation=True)
|
| 26 |
+
|
| 27 |
+
with torch.no_grad():
|
| 28 |
+
# Mixed precision for faster CPU inference
|
| 29 |
+
with torch.cpu.amp.autocast():
|
| 30 |
+
output = self.model.generate(
|
| 31 |
+
**inputs,
|
| 32 |
+
max_new_tokens=300,
|
| 33 |
+
do_sample=False,
|
| 34 |
+
use_cache=True # KV cache speeds up token generation
|
| 35 |
+
)
|
| 36 |
+
|
| 37 |
+
decoded = self.tokenizer.decode(output[0], skip_special_tokens=False)
|
| 38 |
+
decoded = decoded.split("<|im_start|>assistant\n")[1].split("\n<|im_end|>")[0]
|
| 39 |
+
|
| 40 |
+
return decoded
|
| 41 |
+
|
requirements.txt
CHANGED
|
@@ -1,10 +1,7 @@
|
|
| 1 |
fastapi
|
| 2 |
uvicorn
|
| 3 |
-
python-multipart
|
| 4 |
|
| 5 |
-
pandas
|
| 6 |
scikit-learn
|
| 7 |
-
pypdf
|
| 8 |
|
| 9 |
torch --index-url https://download.pytorch.org/whl/cpu
|
| 10 |
|
|
|
|
| 1 |
fastapi
|
| 2 |
uvicorn
|
|
|
|
| 3 |
|
|
|
|
| 4 |
scikit-learn
|
|
|
|
| 5 |
|
| 6 |
torch --index-url https://download.pytorch.org/whl/cpu
|
| 7 |
|