Rework vacancies recommendations

README.md

@@ -18,6 +18,7 @@ uv run --env-file .env manage.py migrate
 uv run --env-file .env manage.py createsuperuser --username stromenko_es --email estromenko@mail.ru
 uv run --env-file .env manage.py runserver
 
+uv run --env-file .env manage.py process_vacancies
 uv run --env-file .env manage.py generate_recommended_vacancies
 uv run --env-file .env manage.py collect_vacancies_from_telegram_messages
 uv run --env-file .env manage.py runbot

pyproject.toml

@@ -10,6 +10,7 @@ dependencies = [
     "langchain>=0.3.27",
     "langchain-openai>=0.3.35",
     "langchain-qdrant>=1.1.0",
+    "pydantic>=2.0",
     "pypdf>=6.1.2",
     "python-telegram-bot>=22.5",
 ]

vacancies/main/features_extractor.py

@@ -0,0 +1,29 @@
+from langchain_openai import ChatOpenAI
+from vacancies.main.models import VacancyFeatures
+
+
+def extract_vacancy_features(content: str) -> VacancyFeatures:
+    """Extract features from vacancy content using structured output."""
+
+    prompt = f"""
+    Extract the following features from the job vacancy description. If a feature is not mentioned, set it to null.
+    Features:
+    - employment_type: Тип занятости (e.g., Полная занятость, Частичная)
+    - work_format: Формат работы (e.g., Офис, Удалённо, Гибрид)
+    - experience: Опыт работы (e.g., 3-5 лет, Нет опыта)
+    - position_level: Уровень позиции (e.g., Junior, Senior)
+    - industry: Отрасль / Сфера деятельности (e.g., IT, Финансы)
+    - tech_stack: Технологический стек / Ключевые навыки (list of strings)
+    - location: География (e.g., Москва, Россия)
+    - salary_range: Зарплатные ожидания / вилка (e.g., 100000-200000 руб)
+    - languages: Языки (list of strings, e.g., ["Русский", "Английский"])
+    - education: Образование (e.g., Высшее, Среднее специальное)
+    - schedule: График работы (e.g., Полный день, Сменный)
+    - additional_requirements: Дополнительные предпочтения / требования (list of strings)
+    Vacancy content:
+    {content}
+    """
+    openai_client = ChatOpenAI(model_name="gpt-5-mini", reasoning_effort="minimal")
+    structured_llm = openai_client.with_structured_output(VacancyFeatures)
+    response = structured_llm.invoke(prompt)
+    return response

vacancies/main/management/commands/collect_vacancies_from_telegram_messages.py

@@ -1,8 +1,8 @@
 from django.core.management import BaseCommand
 from vacancies.main.models import Vacancy
 import clickhouse_connect
-from langchain_core.documents import Document
-from vacancies.main.vector_store import vector_store
+from vacancies.main.vector_store import add_vacancy_vectors
+from vacancies.main.features_extractor import extract_vacancy_features
 
 clickhouse_client = clickhouse_connect.create_client(host="127.0.0.1", port=18123)
 
@@ -25,8 +25,6 @@ class Command(BaseCommand):
     help = "Collect vacancies from telegram messages"
 
     def handle(self, *args, **options):
-        ids, documents = [], []
-
         for index, row in enumerate(clickhouse_client.query(query).result_rows):
             (id, chat_username, telegram_id, message, timestamp) = row
 
@@ -36,10 +34,5 @@ class Command(BaseCommand):
                 defaults={'content': message}
             )
 
-            metadata = {"link": link}
-            ids.append(vacancy.id)
-            documents.append(Document(page_content=message, metadata=metadata))
-
             print(index, link)
-
-        vector_store.add_documents(documents, ids=ids)
+            add_vacancy_vectors(vacancy.id, extract_vacancy_features(message).model_dump(), {"link": link})

vacancies/main/management/commands/generate_recommended_vacancies.py

@@ -1,8 +1,8 @@
 from django.core.management import BaseCommand
-import asyncio
-from vacancies.main.vector_store import vector_store
+from vacancies.main.vector_store import search_similarities
 from vacancies.main.models import CustomerCV, RecommendedVacancy
 from vacancies.main.bot import application
+from vacancies.main.features_extractor import extract_vacancy_features
 from telegram import InlineKeyboardButton, InlineKeyboardMarkup
 from qdrant_client.models import Filter, HasIdCondition
 
@@ -12,7 +12,10 @@ class Command(BaseCommand):
 
     def handle(self, *args, **options):
         customer_cvs = CustomerCV.objects.all()
+
         for customer_cv in customer_cvs:
+            features = extract_vacancy_features(customer_cv.content)
+            print(features)
             recommended_vacancy_ids = RecommendedVacancy.objects.filter(
                 customer=customer_cv.customer
             ).values_list('vacancy_id', flat=True)
@@ -22,22 +25,17 @@ class Command(BaseCommand):
                     HasIdCondition(has_id=recommended_vacancy_ids),
                 ]
             )
+            search_result_id = search_similarities(features.model_dump(), query_filter)
 
-            document = vector_store.similarity_search(
-                customer_cv.content,
-                k=1,
-                filter=query_filter,
-            )[0]
-            print(document.__dict__)
             recommendation, _ = RecommendedVacancy.objects.get_or_create(
                 customer=customer_cv.customer,
-                vacancy_id=document.metadata["_id"],
+                vacancy_id=search_result_id,
             )
 
-            asyncio.run(application.bot.send_message(
+            application.bot.send_message(
                 chat_id=recommendation.customer.chat_id,
                 text=recommendation.vacancy.content,
                 reply_markup=InlineKeyboardMarkup([[
                     InlineKeyboardButton("Откликнуться", url=recommendation.vacancy.link),
                 ]]),
-            ))
+            )

vacancies/main/management/commands/process_vacancies.py

@@ -0,0 +1,20 @@
+from django.core.management import BaseCommand
+from vacancies.main.models import Vacancy
+from vacancies.main.features_extractor import extract_vacancy_features
+from vacancies.main.vector_store import add_vacancy_vectors
+
+
+class Command(BaseCommand):
+    help = "Process vacancies: extract features and index in vector store"
+
+    def handle(self, *args, **options):
+        vacancies = Vacancy.objects.filter(is_processed=False)
+
+        len_vacancies = len(vacancies)
+        for index, vacancy in enumerate(vacancies):
+            print(f"Processing {index}/{len_vacancies} {vacancy}")
+            features = extract_vacancy_features(vacancy.content)
+            vacancy.features_json = features.model_dump()
+            vacancy.is_processed = True
+            vacancy.save()
+            add_vacancy_vectors(vacancy.id, features.model_dump(), {"link": vacancy.link})

vacancies/main/migrations/0004_vacancy_features_json_vacancy_is_processed.py

@@ -0,0 +1,23 @@
+# Generated by Django 5.2.7 on 2025-10-26 11:36
+
+from django.db import migrations, models
+
+
+class Migration(migrations.Migration):
+
+    dependencies = [
+        ('main', '0003_remove_recommendedvacancy_is_shown_and_more'),
+    ]
+
+    operations = [
+        migrations.AddField(
+            model_name='vacancy',
+            name='features_json',
+            field=models.JSONField(blank=True, null=True),
+        ),
+        migrations.AddField(
+            model_name='vacancy',
+            name='is_processed',
+            field=models.BooleanField(default=False),
+        ),
+    ]

vacancies/main/models.py

@@ -1,4 +1,5 @@
 from django.db import models
+from pydantic import BaseModel
 
 
 class Customer(models.Model):
@@ -33,6 +34,8 @@ class CustomerCV(models.Model):
 class Vacancy(models.Model):
     content = models.TextField()
     link = models.URLField(unique=True)
+    features_json = models.JSONField(null=True, blank=True)
+    is_processed = models.BooleanField(default=False)
     created_at = models.DateTimeField(auto_now_add=True)
 
     objects = models.Manager()
@@ -58,3 +61,18 @@ class RecommendedVacancy(models.Model):
     class Meta:
         verbose_name_plural = 'Recommended Vacancies'
         db_table = "recommended_vacancies"
+
+
+class VacancyFeatures(BaseModel):
+    employment_type: str | None = None  # Тип занятости
+    work_format: str | None = None  # Формат работы
+    experience: str | None = None  # Опыт работы
+    position_level: str | None = None  # Уровень позиции
+    industry: str | None = None  # Отрасль / Сфера деятельности
+    tech_stack: list[str] | None = None  # Технологический стек / Ключевые навыки
+    location: str | None = None  # География
+    salary_range: str | None = None  # Зарплатные ожидания / вилка
+    languages: list[str] | None = None  # Языки
+    education: str | None = None  # Образование
+    schedule: str | None = None  # График работы
+    additional_requirements: list[str] | None = None  # Дополнительные предпочтения / требования

vacancies/main/vector_store.py

@@ -1,18 +1,97 @@
-from qdrant_client.models import Distance, VectorParams
-from langchain_qdrant import QdrantVectorStore
+from qdrant_client import models
 from langchain_openai import OpenAIEmbeddings
 from qdrant_client import QdrantClient
+from qdrant_client.models import Filter
 
-client = QdrantClient(path="./embeddings")
+# client = QdrantClient(path="./embeddings")
+client = QdrantClient(url="http://localhost:6333")
 
-if not client.collection_exists("vacancies"):
+FEATURE_NAMES = [
+    "employment_type", "work_format", "experience", "position_level", "industry", "tech_stack",
+    "location", "salary_range", "languages", "education", "schedule", "additional_requirements"
+]
+
+vectors_config = {
+    name: models.VectorParams(size=3072, distance=models.Distance.COSINE) for name in FEATURE_NAMES
+}
+
+collection_name = "vacancies"
+if not client.collection_exists(collection_name):
     client.create_collection(
-        collection_name="vacancies",
-        vectors_config=VectorParams(size=3072, distance=Distance.COSINE)
+        collection_name=collection_name,
+        vectors_config=vectors_config
     )
 
-vector_store = QdrantVectorStore(
-    client=client,
-    collection_name="vacancies",
-    embedding=OpenAIEmbeddings(model="text-embedding-3-large"),
-)
+embedding = OpenAIEmbeddings(model="text-embedding-3-large")
+
+def _prepare_texts(features):
+    """Prepare texts for each feature from features dict."""
+    texts = {}
+    for name in FEATURE_NAMES:
+        value = features.get(name)
+        if isinstance(value, list):
+            text = " ".join(value) if value else ""
+        else:
+            text = str(value) if value else ""
+        texts[name] = text
+    return texts
+
+
+def add_vacancy_vectors(vacancy_id: int, features: dict, payload: dict):
+    """Add vectors for a vacancy based on its features."""
+    texts = _prepare_texts(features)
+    vectors = {}
+    for name, text in texts.items():
+        vectors[name] = [0.0] * 3072
+        if text:
+            vec = embedding.embed_query(text)
+            vectors[name] = vec
+    client.upsert(
+        collection_name=collection_name,
+        points=[
+            models.PointStruct(
+                id=vacancy_id,
+                vector=vectors,
+                payload=payload,
+            )
+        ]
+    )
+
+
+def search_similarities(query_features: dict, query_filter: Filter) -> list[dict]:
+    """Search vacancies using sum of max similarities.
+    For each feature, compute similarities, then for each vacancy, take max per feature, sum.
+    Return top vacancies.
+    """
+    texts = _prepare_texts(query_features)
+    vectors = {}
+    for name, text in texts.items():
+        vectors[name] = [0.0] * 3072
+        if text:
+            vec = embedding.embed_query(text)
+            vectors[name] = vec
+
+    max_similarities = {}
+    for name, vec in vectors.items():
+        if any(v != 0 for v in vec):
+            results = client.search(
+                collection_name=collection_name,
+                query_vector=(name, vec),
+                limit=1000,
+                with_payload=True,
+                query_filter=query_filter,
+            )
+            for res in results:
+                vid = res.id
+                sim = res.score
+                if vid not in max_similarities:
+                    max_similarities[vid] = {}
+                max_similarities[vid][name] = sim
+
+    scored = []
+    for vid, feature_sims in max_similarities.items():
+        total = sum(feature_sims.values())
+        scored.append({"id": vid, "score": total})
+
+    scored.sort(key=lambda x: x["score"], reverse=True)
+    return scored[0]["id"]