Spaces:
Paused
Paused
Commit
·
8d99522
1
Parent(s):
a5be571
updated
Browse files- backend/services/resume_parser.py +44 -49
- requirements.txt +0 -1
backend/services/resume_parser.py
CHANGED
@@ -1,21 +1,18 @@
|
|
1 |
from __future__ import annotations
|
2 |
-
import os
|
3 |
-
import re
|
4 |
-
import subprocess
|
5 |
-
import zipfile
|
6 |
-
import json
|
7 |
-
import torch
|
8 |
from typing import List
|
9 |
-
|
10 |
-
os.environ["OMP_NUM_THREADS"] = "1"
|
11 |
-
os.environ["OPENBLAS_NUM_THREADS"] = "1"
|
12 |
-
os.environ["MKL_NUM_THREADS"] = "1"
|
13 |
-
os.environ["NUMEXPR_NUM_THREADS"] = "1"
|
14 |
-
os.environ["VECLIB_MAXIMUM_THREADS"] = "1"
|
15 |
-
|
16 |
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
|
17 |
-
import torch
|
18 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
bnb_config = BitsAndBytesConfig(
|
20 |
load_in_4bit=True,
|
21 |
bnb_4bit_compute_dtype=torch.float16,
|
@@ -23,10 +20,9 @@ bnb_config = BitsAndBytesConfig(
|
|
23 |
bnb_4bit_quant_type="nf4"
|
24 |
)
|
25 |
|
26 |
-
|
27 |
-
tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/Deepseek-Coder-V2-Lite-Instruct", trust_remote_code=True)
|
28 |
model = AutoModelForCausalLM.from_pretrained(
|
29 |
-
"
|
30 |
quantization_config=bnb_config,
|
31 |
device_map="auto",
|
32 |
torch_dtype=torch.bfloat16,
|
@@ -37,13 +33,10 @@ model = AutoModelForCausalLM.from_pretrained(
|
|
37 |
# Text Extraction (PDF/DOCX)
|
38 |
# ===============================
|
39 |
def extract_text(file_path: str) -> str:
|
40 |
-
"""Extract text from PDF or DOCX resumes."""
|
41 |
if not file_path or not os.path.isfile(file_path):
|
42 |
return ""
|
43 |
-
|
44 |
-
lower_name = file_path.lower()
|
45 |
try:
|
46 |
-
if
|
47 |
result = subprocess.run(
|
48 |
['pdftotext', '-layout', file_path, '-'],
|
49 |
stdout=subprocess.PIPE,
|
@@ -51,8 +44,7 @@ def extract_text(file_path: str) -> str:
|
|
51 |
check=False
|
52 |
)
|
53 |
return result.stdout.decode('utf-8', errors='ignore')
|
54 |
-
|
55 |
-
elif lower_name.endswith('.docx'):
|
56 |
with zipfile.ZipFile(file_path) as zf:
|
57 |
with zf.open('word/document.xml') as docx_xml:
|
58 |
xml_bytes = docx_xml.read()
|
@@ -60,24 +52,20 @@ def extract_text(file_path: str) -> str:
|
|
60 |
xml_text = re.sub(r'<w:p[^>]*>', '\n', xml_text, flags=re.I)
|
61 |
text = re.sub(r'<[^>]+>', ' ', xml_text)
|
62 |
return re.sub(r'\s+', ' ', text)
|
63 |
-
else:
|
64 |
-
return ""
|
65 |
except Exception:
|
66 |
-
|
|
|
67 |
|
68 |
# ===============================
|
69 |
# Name Extraction (Fallback)
|
70 |
# ===============================
|
71 |
def extract_name(text: str, filename: str) -> str:
|
72 |
-
"""Extract candidate's name from resume text or filename."""
|
73 |
if text:
|
74 |
lines = [ln.strip() for ln in text.splitlines() if ln.strip()]
|
75 |
for line in lines[:10]:
|
76 |
-
if re.match(r'(?i)resume|curriculum vitae', line):
|
77 |
-
|
78 |
-
|
79 |
-
if 1 < len(words) <= 4:
|
80 |
-
if all(re.match(r'^[A-ZÀ-ÖØ-Þ][\w\-]*', w) for w in words):
|
81 |
return line
|
82 |
base = os.path.basename(filename)
|
83 |
base = re.sub(r'\.(pdf|docx|doc)$', '', base, flags=re.I)
|
@@ -86,26 +74,25 @@ def extract_name(text: str, filename: str) -> str:
|
|
86 |
return base.title().strip()
|
87 |
|
88 |
# ===============================
|
89 |
-
#
|
90 |
# ===============================
|
91 |
-
def
|
92 |
-
"""Use Deepseek-Coder-V2-Lite-Instruct to extract resume details in JSON format."""
|
93 |
-
|
94 |
prompt = f"""
|
95 |
-
Extract the following information from the resume text provided below.
|
|
|
96 |
|
97 |
Information to extract:
|
98 |
-
- Full Name
|
99 |
-
- Email
|
100 |
-
- Phone
|
101 |
-
- Skills
|
102 |
-
- Education
|
103 |
-
- Experience
|
104 |
|
105 |
-
Resume
|
106 |
{text}
|
107 |
|
108 |
-
|
109 |
{{
|
110 |
"name": "Full Name",
|
111 |
"email": "[email protected]",
|
@@ -115,17 +102,25 @@ Return only valid JSON in the following format:
|
|
115 |
"experience": ["Job1 - Company1 (Dates)", "Job2 - Company2 (Dates)"]
|
116 |
}}
|
117 |
"""
|
118 |
-
|
119 |
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
|
120 |
-
outputs = model.generate(**inputs, max_new_tokens=
|
121 |
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
122 |
|
123 |
-
import re, json
|
124 |
match = re.search(r"\{.*\}", response, re.S)
|
125 |
if match:
|
126 |
try:
|
127 |
return json.loads(match.group())
|
128 |
except:
|
129 |
pass
|
130 |
-
|
131 |
return {"name": "", "email": "", "phone": "", "skills": [], "education": [], "experience": []}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
from __future__ import annotations
|
2 |
+
import os, re, subprocess, zipfile, json, torch
|
|
|
|
|
|
|
|
|
|
|
3 |
from typing import List
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
4 |
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
|
|
|
5 |
|
6 |
+
# Limit threads to avoid Hugging Face Spaces threading issues
|
7 |
+
os.environ.update({
|
8 |
+
"OMP_NUM_THREADS": "1",
|
9 |
+
"OPENBLAS_NUM_THREADS": "1",
|
10 |
+
"MKL_NUM_THREADS": "1",
|
11 |
+
"NUMEXPR_NUM_THREADS": "1",
|
12 |
+
"VECLIB_MAXIMUM_THREADS": "1"
|
13 |
+
})
|
14 |
+
|
15 |
+
# Load Zephyr in 4-bit
|
16 |
bnb_config = BitsAndBytesConfig(
|
17 |
load_in_4bit=True,
|
18 |
bnb_4bit_compute_dtype=torch.float16,
|
|
|
20 |
bnb_4bit_quant_type="nf4"
|
21 |
)
|
22 |
|
23 |
+
tokenizer = AutoTokenizer.from_pretrained("HuggingFaceH4/zephyr-7b-beta", trust_remote_code=True)
|
|
|
24 |
model = AutoModelForCausalLM.from_pretrained(
|
25 |
+
"HuggingFaceH4/zephyr-7b-beta",
|
26 |
quantization_config=bnb_config,
|
27 |
device_map="auto",
|
28 |
torch_dtype=torch.bfloat16,
|
|
|
33 |
# Text Extraction (PDF/DOCX)
|
34 |
# ===============================
|
35 |
def extract_text(file_path: str) -> str:
|
|
|
36 |
if not file_path or not os.path.isfile(file_path):
|
37 |
return ""
|
|
|
|
|
38 |
try:
|
39 |
+
if file_path.lower().endswith('.pdf'):
|
40 |
result = subprocess.run(
|
41 |
['pdftotext', '-layout', file_path, '-'],
|
42 |
stdout=subprocess.PIPE,
|
|
|
44 |
check=False
|
45 |
)
|
46 |
return result.stdout.decode('utf-8', errors='ignore')
|
47 |
+
elif file_path.lower().endswith('.docx'):
|
|
|
48 |
with zipfile.ZipFile(file_path) as zf:
|
49 |
with zf.open('word/document.xml') as docx_xml:
|
50 |
xml_bytes = docx_xml.read()
|
|
|
52 |
xml_text = re.sub(r'<w:p[^>]*>', '\n', xml_text, flags=re.I)
|
53 |
text = re.sub(r'<[^>]+>', ' ', xml_text)
|
54 |
return re.sub(r'\s+', ' ', text)
|
|
|
|
|
55 |
except Exception:
|
56 |
+
pass
|
57 |
+
return ""
|
58 |
|
59 |
# ===============================
|
60 |
# Name Extraction (Fallback)
|
61 |
# ===============================
|
62 |
def extract_name(text: str, filename: str) -> str:
|
|
|
63 |
if text:
|
64 |
lines = [ln.strip() for ln in text.splitlines() if ln.strip()]
|
65 |
for line in lines[:10]:
|
66 |
+
if not re.match(r'(?i)resume|curriculum vitae', line):
|
67 |
+
words = line.split()
|
68 |
+
if 1 < len(words) <= 4 and all(re.match(r'^[A-ZÀ-ÖØ-Þ][\w\-]*', w) for w in words):
|
|
|
|
|
69 |
return line
|
70 |
base = os.path.basename(filename)
|
71 |
base = re.sub(r'\.(pdf|docx|doc)$', '', base, flags=re.I)
|
|
|
74 |
return base.title().strip()
|
75 |
|
76 |
# ===============================
|
77 |
+
# Zephyr Parsing
|
78 |
# ===============================
|
79 |
+
def parse_with_zephyr(text: str) -> dict:
|
|
|
|
|
80 |
prompt = f"""
|
81 |
+
Extract the following information from the resume text provided below.
|
82 |
+
Return ONLY a valid JSON object (no extra commentary).
|
83 |
|
84 |
Information to extract:
|
85 |
+
- Full Name
|
86 |
+
- Email
|
87 |
+
- Phone
|
88 |
+
- Skills (list)
|
89 |
+
- Education (list of degrees + institutions)
|
90 |
+
- Experience (list of jobs with company, title, and dates)
|
91 |
|
92 |
+
Resume:
|
93 |
{text}
|
94 |
|
95 |
+
JSON format:
|
96 |
{{
|
97 |
"name": "Full Name",
|
98 |
"email": "[email protected]",
|
|
|
102 |
"experience": ["Job1 - Company1 (Dates)", "Job2 - Company2 (Dates)"]
|
103 |
}}
|
104 |
"""
|
|
|
105 |
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
|
106 |
+
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.0)
|
107 |
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
108 |
|
|
|
109 |
match = re.search(r"\{.*\}", response, re.S)
|
110 |
if match:
|
111 |
try:
|
112 |
return json.loads(match.group())
|
113 |
except:
|
114 |
pass
|
|
|
115 |
return {"name": "", "email": "", "phone": "", "skills": [], "education": [], "experience": []}
|
116 |
+
|
117 |
+
# ===============================
|
118 |
+
# Main Parse Function
|
119 |
+
# ===============================
|
120 |
+
def parse_resume(file_path: str, filename: str) -> dict:
|
121 |
+
text = extract_text(file_path)
|
122 |
+
name_fallback = extract_name(text, filename)
|
123 |
+
data = parse_with_zephyr(text)
|
124 |
+
if not data.get("name"):
|
125 |
+
data["name"] = name_fallback
|
126 |
+
return data
|
requirements.txt
CHANGED
@@ -62,5 +62,4 @@ requests>=2.31.0
|
|
62 |
psycopg2-binary
|
63 |
matplotlib
|
64 |
bitsandbytes>=0.41.0
|
65 |
-
flash-attn==2.3.6 --no-build-isolation
|
66 |
|
|
|
62 |
psycopg2-binary
|
63 |
matplotlib
|
64 |
bitsandbytes>=0.41.0
|
|
|
65 |
|