init

.gitattributes

@@ -33,3 +33,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.csv filter=lfs diff=lfs merge=lfs -text
+*.sqlite3 filter=lfs diff=lfs merge=lfs -text
+*.rdb filter=lfs diff=lfs merge=lfs -text
+*.json filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,5 @@
+__pycache__/
+app-data/
+redis-data/
+k8s/
+*.rdb

Dockerfile

@@ -0,0 +1,23 @@
+FROM python:3.9-slim
+
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    build-essential \
+    && rm -rf /var/lib/apt/lists/*
+
+
+COPY requirements.txt .
+RUN pip install --no-cache-dir -r requirements.txt
+
+
+RUN export PYTHONUNBUFFERED=1
+RUN useradd -m -u 1000 user
+USER user
+
+WORKDIR /app
+COPY app /app
+
+
+EXPOSE 7860
+
+CMD ["python", "-u", "app.py"]

README copy.md

@@ -0,0 +1,10 @@
+---
+title: Course Rec
+emoji: 🏢
+colorFrom: purple
+colorTo: green
+sdk: docker
+pinned: false
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app/app.py

@@ -0,0 +1,235 @@
+import gradio as gr
+import pandas as pd
+import redis
+import json
+import requests
+from config import *
+import functools
+from embedding_setup import retriever, find_similar_occupation, compare_docs_with_context,generate_exp,generate_prompt_exp
+from data_process import  get_occupations_from_csv, get_courses_from_BA, get_occupation_detial, build_occupation_query
+with open('/app/data/redis_data.json', 'r') as file:
+    data_dict = json.load(file)
+#r = redis.Redis(host=REDIS_HOST, port=REDIS_PORT, db=REDIS_DB, decode_responses=True)
+
+skill_details_mapping = {}
+
+
+# Function to retrieve documents based on selected skills
+def retrieve_documents(occupation,skills):
+    output = []
+    output.append(f"<div style=\"text-align: center; font-size: 24px;\">Empfehlungsergebnisse:</div>")
+    oc_uri = occupations.get(occupation, "")
+    skill_query = ''
+    candidate_docs = []
+    if isinstance(oc_uri, int):
+        df = pd.read_csv("/app/data/berufe_info.csv")
+        target_occupation = df[df['id'] == oc_uri]
+        target_occupation_name = target_occupation['short name'].values[0]
+        target_occupation_dsp = target_occupation['description'].values[0]
+        target_occupation_query = target_occupation_name + ' ' + target_occupation_dsp
+        target_occupation_query = target_occupation_query
+    else:
+        target_occupation = get_occupation_detial(oc_uri)
+        target_occupation_name, target_occupation_dsp, target_occupation_query = build_occupation_query(target_occupation)
+    for german_label in skills:
+        skill_query += german_label + ' '
+        ocsk_query = target_occupation_name + ' ' + german_label
+        skills_docs = retriever.get_relevant_documents(ocsk_query)
+        candidate_docs.extend(skills_docs[:2])
+    query =  target_occupation_query + ' ' + skill_query
+    llama_query = 'info:' + target_occupation_name + ' ' + 'Skills gap:' + skill_query
+    print(query)
+    docs = retriever.get_relevant_documents(query)
+    candidate_docs.extend(docs[:5])
+
+    #remove duplicates
+    seen_course_ids = set()
+    candidate_doc_unique = []
+
+    for doc in candidate_docs:
+        course_id = doc.metadata.get('id','')
+        if course_id not in seen_course_ids:
+            candidate_doc_unique.append(doc)
+            seen_course_ids.add(course_id)
+
+    partial_compare_docs = functools.partial(compare_docs_with_context, target_occupation_name=target_occupation_name, target_occupation_dsp=target_occupation_dsp,skill_gap = skill_query)
+    sorted_docs = sorted(candidate_doc_unique, key=functools.cmp_to_key(partial_compare_docs), reverse=True)
+
+    
+    batch_output = []
+    for doc in sorted_docs[:5]:
+        doc_name = doc.metadata.get('name', 'Unnamed Document')
+        doc_skill = doc.metadata.get('skills', '')
+        input_text = f"target occupation: {llama_query}\n Recommended course: name: {doc_name}, learning objectives: {doc_skill[:2000]}"
+        prompt = generate_prompt_exp(input_text)
+        batch_output += generate_exp(prompt)
+
+    # Evaluate the current batch of prompts
+    output.append(f"<b>Zielberuf:</b> {target_occupation_name}")
+    output.append(f"<b>Qualifikationslücke:</b> {skill_query}")
+    output.append(f"<b>Empfohlene Kurse:</b>")
+    for i in range(5):
+        doc = sorted_docs[i]
+        doc_name = doc.metadata.get('name', 'Unnamed Document')
+        doc_url = doc.metadata.get('url', '#')
+        doc_skill = doc.metadata.get('skills', '')
+        output.append(f"<a href='{doc_url}' target='_blank'>{doc_name}</a>") 
+        output.append(f"<b>Empfehlungsgrund:</b> {batch_output[i]}")
+    
+
+    output.append(f"<br>")
+    return "<br>".join(output)
+
+
+def get_candidate_courses(occupation, skills):
+    output = []
+    output.append(f"<div style=\"text-align: center; font-size: 24px;\">Empfehlungsergebnisse:</div>")
+    df_lookup = pd.read_csv('/app/data/kldb_isco_lookup.csv')
+    df_berufe = pd.read_csv('/app/data/berufe_info.csv')
+    occupation_codes = set()
+    kldB_set = set()
+    occupation_hrefs = set()
+    BA_berufe = set()
+    oc_uri = occupations.get(occupation, "")
+    target_occupation = get_occupation_detial(oc_uri)
+    target_occupation_query = build_occupation_query(target_occupation)
+    
+    for german_label in skills:
+        skill = skill_details_mapping.get(german_label, {})
+        uri = f'https://ec.europa.eu/esco/api/resource/skill?selectedVersion=v1.0.9&language=en&uri={skill["uri"]}'
+        try:
+            skill_response = requests.get(uri)
+            skill_response.raise_for_status()
+            skill_json = skill_response.json()
+            
+            # Combine essential and optional occupations
+            skill_related_occupations = (skill_json['_links'].get('isEssentialForOccupation', []) +
+                                          skill_json['_links'].get('isOptionalForOccupation', []))
+            
+            for occupation in skill_related_occupations:
+                href = occupation.get('href')
+                if href:
+                    occupation_hrefs.add(href)
+        except requests.RequestException as e:
+            print(f"Error while fetching skill details: {e}")
+                
+    for href in occupation_hrefs:
+        try:
+            occupation_response = requests.get(href)
+            occupation_response.raise_for_status()
+            occupation_details = occupation_response.json()
+            
+            code = occupation_details.get('code')
+            if code:
+                occupation_codes.add(code.split('.')[0])
+        except requests.RequestException as e:
+            print(f"Error while fetching occupation details: {e}")
+            
+    for isco_code in occupation_codes:
+        kldB_codes = df_lookup[df_lookup['isco08'] == int(isco_code)]['kldb2010'].values
+        for code in kldB_codes:
+            kldB_set.add(str(code))
+    dfs = []
+    for kldb in kldB_set:
+        berufe = df_berufe[df_berufe['KldB codes']=='B '+kldb]
+        dfs.append(berufe)
+
+    merged_df = pd.concat(dfs, ignore_index=True)  
+    top_k_berufe = find_similar_occupation(target_occupation_query,merged_df,5,'cosine')
+    for beruf in top_k_berufe:
+        entry_requirement = beruf.metadata['entry_requirements']
+        corrected_json_string = entry_requirement.replace("'", '"')
+        entry_requirement_json = json.loads(corrected_json_string)
+        for js in entry_requirement_json:
+            BA_berufe.add(str(js['data_idref']))
+                
+    result = get_courses_from_BA(BA_berufe)
+    courses = result
+    for course in courses['_embedded']['termine']:
+        output.append(f"<a href='{course['angebot']['link']}' target='_blank'>{course['angebot']['titel']}</a>") 
+
+    return "<br>".join(output)
+
+
+def get_occupation_skills(oc_uri):
+    #skills_json = r.get(oc_uri)
+    skills_json = data_dict.get(oc_uri, None)
+    skill_labels = []
+    if skills_json:
+        skills = json.loads(skills_json)
+        for skill in skills:
+            german_label = skill['preferredLabel']['de']
+            skill_details_mapping[german_label] = skill
+            skill_labels.append(german_label)
+        return skill_labels
+    else:
+        return skill_labels
+    
+def get_occupation_skills_BA(oc_uri):
+    df = pd.read_csv("/app/data/berufe_info.csv")
+    essential_skills = df[df['id'] == oc_uri]['essential skills'].values
+    optional_skills = df[df['id'] == oc_uri]['optional skills'].values
+    combined_skills = essential_skills[0][:-1] + ',' + optional_skills[0][1:]
+    combined_skills = combined_skills.replace("'", "\"")
+    skills = json.loads(combined_skills)
+    skill_labels = []
+    for skill in skills:
+        german_label = skill['skill']
+        skill_details_mapping[german_label] = skill
+        skill_labels.append(german_label)
+    return skill_labels
+
+# Function to update the skills dropdown
+def update_skills(occupation):
+    oc_uri = occupations.get(occupation, "")
+    if isinstance(oc_uri, int):
+        skills = get_occupation_skills_BA(oc_uri)
+        return gr.Dropdown(skills,label="aktuelle Fähigkeiten", multiselect=True,info='Bitte wählen Sie die Fähigkeiten aus, die Sie derzeit besitzen')
+    else:
+        skills = get_occupation_skills(oc_uri)
+        return gr.Dropdown(skills,label="aktuelle Fähigkeiten", multiselect=True,info='Bitte wählen Sie die Fähigkeiten aus, die Sie derzeit besitzen')
+    return 
+
+def update_skillgap(occupation, current_skills):
+    oc_uri = occupations.get(occupation, "")
+    if isinstance(oc_uri, int):
+        ocupation_skills = get_occupation_skills_BA(oc_uri)
+    else:
+        ocupation_skills = get_occupation_skills(oc_uri)
+    skill_gap = [skill for skill in ocupation_skills if skill not in current_skills]
+    
+    return gr.Dropdown(skill_gap, label="Qualifikationslücke", multiselect=True, info='Bitte wählen Sie die Fähigkeiten aus, die Sie lernen möchten.')
+
+if __name__ == "__main__":
+    # Load occupations from CSV
+    occupations_esco = get_occupations_from_csv(CSV_FILE_PATH)
+    df = pd.read_csv("/app/data/berufe_info.csv")
+    occupations_BA = df[['short name', 'id']].set_index('short name').to_dict()['id']
+    occupations = {**occupations_esco, **occupations_BA}
+    # Gradio interface
+    with gr.Blocks(title="MyEduLife Kursempfehlungssystem") as demo:
+        occupation_dropdown = gr.Dropdown(list(occupations.keys()), label="Zielberuf",info='Bitte wählen Sie Ihren Zielberuf aus.')
+        currentskill_dropdown = gr.Dropdown([],label="aktuelle Fähigkeiten", multiselect=True,info='Bitte wählen Sie die Fähigkeiten aus, die Sie derzeit besitzen')
+        sb_btn = gr.Button("Absenden")
+        skillgap_dropdown = gr.Dropdown([],label="Fähigkeiten", multiselect=True,info='Bitte wählen Sie die Fähigkeiten aus, die Sie lernen möchten.')
+                # Use gr.HTML to display the HTML content
+        button = gr.Button("Kursempfehlungen")
+        documents_output = gr.HTML()
+
+        occupation_dropdown.change(update_skills, inputs=occupation_dropdown, outputs=currentskill_dropdown)
+
+        sb_btn.click(
+                    update_skillgap, 
+                    inputs=[occupation_dropdown,currentskill_dropdown], 
+                    outputs=skillgap_dropdown
+                )
+
+        button.click(
+                    retrieve_documents, 
+                    inputs=[occupation_dropdown,skillgap_dropdown],
+                    outputs=documents_output
+                    )
+    print('Initialization completed')
+    demo.launch(server_name="0.0.0.0", server_port=7860)
+    
+    

app/config.py

@@ -0,0 +1,26 @@
+import os
+
+# Redis configuration
+#REDIS_HOST = os.getenv('REDIS_HOST', 'redis')
+REDIS_HOST = os.getenv('REDIS_HOST', '0.0.0.0')
+REDIS_PORT = int(os.getenv('REDIS_PORT', 6379))  
+REDIS_DB = int(os.getenv('REDIS_DB', 0))
+
+# Model and embedding configuration
+#MODEL_NAME = os.getenv('MODEL_NAME', "intfloat/multilingual-e5-large-instruct")
+MODEL_NAME = os.getenv('MODEL_NAME', "wt3639/EduGBERT_CourseRec")
+ENCODE_KWARGS = {
+    'normalize_embeddings': os.getenv('NORMALIZE_EMBEDDINGS', 'True') == 'True',  
+    'convert_to_tensor': os.getenv('CONVERT_TO_TENSOR', 'True') == 'True'
+}
+#QUERY_INSTRUCTION = os.getenv('QUERY_INSTRUCTION', 'Find the course that relates to the given occupation and cover the skills gap')
+QUERY_INSTRUCTION = os.getenv('QUERY_INSTRUCTION', '')
+# Other configurations
+TOP_K = int(os.getenv('TOP_K', 10))
+#PERSIST_DIRECTORY = os.getenv('PERSIST_DIRECTORY', "/app/data/course_emb_db")
+PERSIST_DIRECTORY = os.getenv('PERSIST_DIRECTORY', "/app/data/EduGBERT_cos_escoai")
+CSV_FILE_PATH = os.getenv('CSV_FILE_PATH', '/app/data/occupations_de.csv')
+
+REC_LORA_MODEL = os.getenv('REC_LORA_MODEL', 'wt3639/Llama-3-8B-Instruct_CourseRec_lora')
+EXP_LORA_MODEL = os.getenv('EXP_LORA_MODEL', 'wt3639/Lllama-3-8B-instruct-exp-adapter')
+LLM_MODEL = os.getenv('LLM_MODEL', 'meta-llama/Meta-Llama-3-8B-Instruct')

app/data/EduGBERT_cos_escoai/chroma.sqlite3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6a66289aa47eb8a5339cb507db5998aaf0027b7638ee6750d26a1459e7b64a9a
+size 103567360

app/data/EduGBERT_cos_escoai/facbf936-cb42-45b3-8dd7-c6a84203b62a/data_level0.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bfcaa4e2f82fc0426cfa574fc0f5126d913d426925e1ac3cdd3403b0d6d90d57
+size 25416000

app/data/EduGBERT_cos_escoai/facbf936-cb42-45b3-8dd7-c6a84203b62a/header.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7ffd5ef9bb9ca1a421e18216d2997eea01192524736b484b49cd43b46e42e31f
+size 100

app/data/EduGBERT_cos_escoai/facbf936-cb42-45b3-8dd7-c6a84203b62a/index_metadata.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:56f054e8946b4b38b357da496aca2bebffc0adf8010a12a1bb4218581d1bf8ca
+size 346049

app/data/EduGBERT_cos_escoai/facbf936-cb42-45b3-8dd7-c6a84203b62a/length.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b3a5a47864c30653e6f7a237ebf2b02ede483e75377c439e672417c42c0c6e0
+size 24000

app/data/EduGBERT_cos_escoai/facbf936-cb42-45b3-8dd7-c6a84203b62a/link_lists.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1c637f61ab8c840f644f6a5d9c6a6290a5374997e1d3973225841d93c3287e2b
+size 51880

app/data/ba_course_emb/3b822eb4-09a8-4577-a432-58654a23254b/data_level0.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:47834eba6ccc20700681289c6198c4c2f6824715763ad8be26953ba7748194c2
+size 4236000

app/data/ba_course_emb/3b822eb4-09a8-4577-a432-58654a23254b/header.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f03662fe47efebf7a7d3d9b8b5341473d199915b17d7cf59ed78ea2bf986a168
+size 100

app/data/ba_course_emb/3b822eb4-09a8-4577-a432-58654a23254b/index_metadata.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48c1e4f32bde583bd6a0a03cf194ff0021751797d9ef32b210427f3e75437c2a
+size 55974

app/data/ba_course_emb/3b822eb4-09a8-4577-a432-58654a23254b/length.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48bff893dbcb9255890ac7dc0d30b26e4f781f6ae5c77debb693fd04c68c2d79
+size 4000

app/data/ba_course_emb/3b822eb4-09a8-4577-a432-58654a23254b/link_lists.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b01c1d32064620009fd560852b48bdb9eebd0ba3050b1b2ca553918602715771
+size 8420

app/data/ba_course_emb/chroma.sqlite3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:166d64d9df06a4494ad9fff6e9b37f8df33f710514d69c16b9fc1984d6066b6d
+size 38019072

app/data/berufe_info.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa16cd91e4ae94d8044b4421aa2392b4918245896745677ba471770eeb0f69bc
+size 11483882

app/data/kldb_isco_lookup.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6918ef9566992565deb01b1ca79b31b50c21130e2dcd42b9a21e116b7ddaa7a9
+size 363652

app/data/multilingual-e5_cos_escoai/c4b78092-a1fa-41fc-b6b0-8fa36ef29576/data_level0.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5febe5913f14ef4735dcb9551b6f86e9d183b2fac0a2c1ac980fee33a6443534
+size 25416000

app/data/multilingual-e5_cos_escoai/c4b78092-a1fa-41fc-b6b0-8fa36ef29576/header.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7ffd5ef9bb9ca1a421e18216d2997eea01192524736b484b49cd43b46e42e31f
+size 100

app/data/multilingual-e5_cos_escoai/c4b78092-a1fa-41fc-b6b0-8fa36ef29576/index_metadata.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f91c55d437d031ea4094c2a2c3e35e0aec42eb5d5f60821d7b164f714228dbe
+size 346049

app/data/multilingual-e5_cos_escoai/c4b78092-a1fa-41fc-b6b0-8fa36ef29576/length.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1d3b0223a8c3ef4338003b319ca7fe73feb39b300f9dc618e6a9373083b59f6a
+size 24000

app/data/multilingual-e5_cos_escoai/c4b78092-a1fa-41fc-b6b0-8fa36ef29576/link_lists.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95c0583c98e9b4f9d1694015eb8a2bd782fca6f6ad14d7580ba83b1ced231e26
+size 51880

app/data/multilingual-e5_cos_escoai/chroma.sqlite3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e5e1bfab026e95b44810b2284b3f82d10871fcd4242aa122e31fa84e27195209
+size 103567360

app/data/occupations_de.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4734be24b6ce55415c5e872dbe91134541189935c776765a4ef96aa35db4b282
+size 2478772

app/data/redis_data.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a01eb57f05a046beff74f8bdf2126a24a3996c0ffdeaaa444c423aaa40ee2e4
+size 117289041

app/data_process.py

@@ -0,0 +1,123 @@
+from config import *
+import pandas as pd
+import requests
+
+
+
+
+def build_skill_query(skill):
+    skill_query =  skill['preferredLabel']['de'] +" " + skill['preferredLabel']['en']+" "+ skill['description']['de']+ " "+ skill['description']['en']
+    if skill['borderConcept']['broaderHierarchyConcept']:
+        broader_hierarchy_concept_str = ", ".join(skill['borderConcept']['broaderHierarchyConcept'])
+        skill_query += " " + broader_hierarchy_concept_str
+    else:
+        pass
+    if skill['borderConcept']['broaderSkill']:
+        broaderSkill_str = ", ".join(skill['borderConcept']['broaderSkill'])
+        skill_query += " " + broaderSkill_str
+    else:
+        pass
+    if skill['alternativeLabel']['de']:
+        alternativeLabel_de_str = ", ".join(skill['alternativeLabel']['de'])
+        skill_query += " " + alternativeLabel_de_str
+    else:
+        pass
+    if skill['alternativeLabel']['en']:
+        alternativeLabel_en_str = ", ".join(skill['alternativeLabel']['en'])
+        skill_query += " " + alternativeLabel_en_str
+    else:
+        pass 
+    return skill_query
+
+
+
+def build_occupation_query(occupation):
+    occupation_name_de = occupation['preferredLabel'].get('de','')
+    occupation_dsp = occupation['description'].get('de','').get('literal','')
+    occupation_query = occupation_name_de +" " + occupation['preferredLabel'].get('en','')+" "+ occupation['description'].get('de','').get('literal','') + " "+ occupation_dsp
+    '''
+    if occupation['_links']['broaderIscoGroup']:
+        for group in occupation['_links']['broaderIscoGroup']:
+            occupation_query += " " + group['title']
+    else:
+        pass
+        '''
+    return occupation_name_de,occupation_dsp,occupation_query      
+
+# Get occupations from a CSV
+def get_occupations_from_csv(file_path):
+    df = pd.read_csv(file_path)
+    return df[['preferredLabel', 'conceptUri']].set_index('preferredLabel').to_dict()['conceptUri']
+
+
+def get_oauth_token():
+    # API endpoint URL
+    token_url = "https://rest.arbeitsagentur.de/oauth/gettoken_cc"
+    
+    # Client credentials
+    client_id = "38053956-6618-4953-b670-b4ae7a2360b1"
+    client_secret = "c385073c-3b97-42a9-b916-08fd8a5d1795"
+    grant_type = "client_credentials"
+    
+    # Prepare request data
+    payload = {
+        "client_id": client_id,
+        "client_secret": client_secret,
+        "grant_type": grant_type
+    }
+    
+    # Send request and get response
+    response = requests.post(token_url, data=payload)
+    
+    # Check if the request was successful
+    if response.status_code == 200:
+        return response.json().get("access_token")
+    else:
+        print("Token request failed:", response.text)
+        return None
+
+def query_weiterbildungssuche_api(token, params):
+    # Set API URL
+    api_url = "https://rest.arbeitsagentur.de/infosysbub/wbsuche/pc/v2/bildungsangebot"
+    
+    # Prepare request headers
+    headers = {
+        "Authorization": f"Bearer {token}"
+    }
+    
+    # Send GET request
+    response = requests.get(api_url, headers=headers, params=params)
+    
+    if response.status_code == 200:
+        return response.json()
+    else:
+        print("API request failed:", response.text)
+        return None
+
+def get_courses_from_BA(ids):
+    # Get OAuth token
+    token = get_oauth_token()
+    if token:
+        # Set query parameters
+        params = {
+            "ids": list(ids)
+        }
+        # Use token to query the API
+        result = query_weiterbildungssuche_api(token, params)
+        return result
+
+
+def get_occupation_detial(oc_uri):
+    uri = f'https://ec.europa.eu/esco/api/resource/occupation?selectedVersion=v1.0.9&language=en&uri={oc_uri}'
+    try:
+            occupation_response = requests.get(uri)
+            occupation_response.raise_for_status()
+            occupation_json = occupation_response.json()
+            return occupation_json
+            
+    except requests.RequestException as e:
+        print(f"Error while fetching skill details: {e}")
+
+
+
+

app/embedding_setup.py

@@ -0,0 +1,204 @@
+from langchain_community.vectorstores import Chroma
+from langchain_community.embeddings import HuggingFaceBgeEmbeddings
+
+from langchain.docstore.document import Document
+from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig,BitsAndBytesConfig
+from peft import PeftModel
+from config import *
+import os
+import torch
+from predibase import Predibase, FinetuningConfig, DeploymentConfig
+pb_auth  = os.environ.get("pb_token")
+pb = Predibase(api_token=pb_auth)
+
+if torch.cuda.is_available():
+    device = "cuda"
+else:
+    device = "cpu"
+
+os.environ['CURL_CA_BUNDLE'] = ""
+embedding_int = HuggingFaceBgeEmbeddings(
+    model_name=MODEL_NAME,
+    encode_kwargs=ENCODE_KWARGS,
+    query_instruction=QUERY_INSTRUCTION
+)
+
+embedding_sim = HuggingFaceBgeEmbeddings(
+    model_name=MODEL_NAME,
+    encode_kwargs=ENCODE_KWARGS,
+    query_instruction='Retrieve semantically similar text.'
+)
+
+db = Chroma(persist_directory=PERSIST_DIRECTORY, embedding_function=embedding_int)
+retriever = db.as_retriever(search_kwargs={"k": TOP_K})
+
+
+
+lora_weights_rec = REC_LORA_MODEL
+lora_weights_exp = EXP_LORA_MODEL
+hf_auth  = os.environ.get("hf_token")
+
+tokenizer = AutoTokenizer.from_pretrained(LLM_MODEL, token=hf_auth)
+
+
+first_token = 'First'
+second_token = 'Second'
+# 获取token的ID
+first_id = tokenizer.convert_tokens_to_ids(first_token)
+second_id = tokenizer.convert_tokens_to_ids(second_token)
+model = AutoModelForCausalLM.from_pretrained(
+            LLM_MODEL,
+            load_in_4bit=True,
+            torch_dtype=torch.float16,
+            token=hf_auth,
+        )
+
+rec_adapter = PeftModel.from_pretrained(
+            model,
+            lora_weights_rec
+        )
+      
+
+tokenizer.padding_side = "left"
+    # unwind broken decapoda-research config
+#model.half()  # seems to fix bugs for some users.
+rec_adapter.eval()
+
+rec_adapter.config.pad_token_id = tokenizer.pad_token_id = 0  # unk
+rec_adapter.config.bos_token_id = 1
+rec_adapter.config.eos_token_id = 2
+
+
+
+def generate_prompt(target_occupation, skill_gap, courses):
+    return f"""
+### Instruction:
+"As an education expert, you have been provided with a target occupation, a skill gap, and information on two candidate courses. Your task is to determine which course better matches the target occupation and skill gap. Please respond with 'First' or 'Second' to indicate your recommendation.
+
+### Input:
+Target Occupation: {target_occupation}
+Skill Gap: {skill_gap}
+candidate courses: {courses}
+
+### Response:
+"""
+'''
+prompt_re = ChatPromptTemplate.from_template(template_re)
+chain_re = (
+    runnable
+    | prompt_re
+)
+'''
+def evaluate(
+        prompt=None,
+        temperature=0,
+        top_p=1.0,
+        top_k=40,
+        num_beams=1,
+        max_new_tokens=30,
+        batch_size=1,
+        **kwargs,
+    ):
+
+        inputs = tokenizer(prompt, return_tensors="pt", padding=True, truncation=True).to(device)
+        generation_config = GenerationConfig(
+            temperature=temperature,
+            top_p=top_p,
+            top_k=top_k,
+            num_beams=num_beams,
+            **kwargs,
+        )
+        with torch.no_grad():
+            rec_adapter.to(device)
+            generation_output = rec_adapter.generate(
+                **inputs,
+                generation_config=generation_config,
+                return_dict_in_generate=True,
+                output_scores=True,
+                max_new_tokens=max_new_tokens,
+                # batch_size=batch_size,
+                eos_token_id=tokenizer.eos_token_id,
+                pad_token_id=tokenizer.eos_token_id,
+            )
+        scores = generation_output.scores[0].softmax(dim=-1)
+        logits = torch.tensor(scores[:,[first_id, second_id]], dtype=torch.float32).softmax(dim=-1)
+        s = generation_output.sequences
+        output = tokenizer.batch_decode(s, skip_special_tokens=True)
+        output = [_.split('Response:\n')[-1] for _ in output]
+        return output, logits.tolist()
+
+def compare_docs_with_context(doc_a, doc_b, target_occupation_name, target_occupation_dsp,skill_gap):
+    
+    #courses = f"First: name: {doc_a.metadata['name']}  description:{doc_a.metadata['description']} Second: name: {doc_b.metadata['name']}  description:{Sdoc_b.metadata['description']}" 
+    courses = f"First: name: {doc_a.metadata['name']}  learning outcomes:{doc_a.metadata['skills'][:1500]} Second: name: {doc_b.metadata['name']}  learning outcomes:{doc_b.metadata['skills'][:1500]}" 
+    target_occupation = f"name: {target_occupation_name} description: {target_occupation_dsp[:1500]}"
+    skill_gap = skill_gap
+    prompt = generate_prompt(target_occupation, skill_gap, courses)
+    prompt = [prompt]
+    output, logit = evaluate(prompt)
+    # Compare based on the response: [A] means doc_a > doc_b, [B] means doc_a < doc_b
+    print(output, logit)
+    if logit[0][0] > logit[0][1]:
+        return 1  # doc_a should come before doc_b
+    elif logit[0][0] < logit[0][1]:
+        return -1  # doc_a should come after doc_b
+    else:
+        return 0  # Consider them equal if the response is unclear
+
+
+#-----------------------------------------explanation-------------------------------------
+
+lorax_client = pb.deployments.client("llama-3-8b-instruct") # Insert deployment name here
+def generate_prompt_exp(input_text):
+    return f"""
+### Instruction:
+As an education expert, you have been provided with information on target occupations and skills gaps, along with recommended course details. Your task is to explain the recommendation in German, focusing on how the course's learning outcomes and target skills relate to the identified skills gaps.
+
+### Input:
+{input_text}
+
+### Response:
+"""
+
+def generate_exp(
+        prompt=None,
+        temperature=0.2,
+        top_p=1.0,
+        top_k=40,
+        num_beams=1,
+        max_new_tokens=512,
+        batch_size=1,
+        do_sample=True,
+        **kwargs,
+    ):
+
+        
+        resp = lorax_client.generate(prompt,adapter_id="wt3639/Llama-3-8B-Instruct_RecExp_lora", adapter_source='hub', max_new_tokens=max_new_tokens, temperature=temperature, top_p=top_p, top_k=top_k, do_sample=do_sample)
+        
+        return resp.generated_text
+
+
+def find_similar_occupation(target_occupation_query, berufe, top_k, similarity_func):
+  
+    # Pro kurs wird ein Document erstellt. Dieses enthält Metadaten sowie einen page_content. 
+    # Der Inhalt von page_content wird embedded und so für die sucher verwendet.
+    docs = []
+    for index, beruf in berufe.iterrows():  
+        # Create document.
+        doc = Document(
+            page_content= beruf['short name'] + ' ' + beruf['full name'] + ' ' + beruf['description'],  
+            metadata={
+                "id": beruf["id"],
+                "name": beruf['short name'],
+                "description": beruf["description"],
+                "entry_requirements": beruf["entry requirements"]
+            },
+        )
+        docs.append(doc)
+    
+    db_temp = Chroma.from_documents(documents = docs, embedding= embedding_sim, collection_metadata = {"hnsw:space": similarity_func})
+    # Retriever will search for the top_5 most similar documents to the query.
+    retriever_temp = db_temp.as_retriever(search_kwargs={"k": top_k})
+    top_similar_occupations = retriever_temp.get_relevant_documents(target_occupation_query)
+
+    return top_similar_occupations