Optimize vacancies vectorization

vacancies/main/bot.py

@@ -1,15 +1,34 @@
-import os
 import io
+import os
 import traceback
-from telegram import InlineKeyboardButton, InlineKeyboardMarkup, Update, ReplyKeyboardMarkup, KeyboardButton
-from telegram.ext import filters, ApplicationBuilder, MessageHandler, CommandHandler, ContextTypes
-from pypdf import PdfReader
-from vacancies.main.models import Customer, CustomerCV
+
 from langchain.agents import create_agent
 from langchain_openai import ChatOpenAI
 from langgraph.checkpoint.postgres.aio import AsyncPostgresSaver
-from vacancies.main.vector_store import add_vectors, extract_features, get_next_vacancy
+from pypdf import PdfReader
+from telegram import (
+    InlineKeyboardButton,
+    InlineKeyboardMarkup,
+    KeyboardButton,
+    ReplyKeyboardMarkup,
+    Update,
+)
+from telegram.ext import (
+    ApplicationBuilder,
+    CommandHandler,
+    ContextTypes,
+    MessageHandler,
+    filters,
+)
+
 from vacancies.conf.settings import DB_URI
+from vacancies.main.models import Customer, CustomerCV
+from vacancies.main.vector_store import (
+    add_vectors,
+    batch_extract_features,
+    get_next_vacancy,
+    embed_features,
+)
 
 SYSTEM_PROMPT = """
 Ты — карьерный копилот для ИТ. Ты можешь отвечать на любые вопросы по тематике карьеры.
@@ -107,11 +126,11 @@ async def handle_document(update: Update, context: ContextTypes.DEFAULT_TYPE):
     customer_cv, _ = await CustomerCV.objects.aupdate_or_create(customer=customer, defaults=dict(
         content=resume,
     ))
-    features = extract_features(customer_cv.content)
+    features = batch_extract_features(customer_cv.content)[0]
     add_vectors(
         "cvs",
         customer_cv.id,
-        features.model_dump(),
+        embed_features(features.model_dump())[0],
         {'content': customer_cv.content, 'features_json': features.model_dump()},
     )
 

vacancies/main/management/commands/collect_vacancies_from_telegram_messages.py

@@ -1,19 +1,23 @@
-import traceback
-from itertools import batched
 from concurrent.futures import ThreadPoolExecutor
+from datetime import datetime, timedelta
+from itertools import batched
 
+import clickhouse_connect
 from django.core.management import BaseCommand
 from django.conf import settings
-import clickhouse_connect
-from vacancies.main.vector_store import add_vectors, extract_features, qdrant_client
-from vacancies.conf.settings import CLICKHOUSE_HOST, CLICKHOUSE_PORT
+from qdrant_client.models import OrderBy
 
-clickhouse_client = clickhouse_connect.create_client(host=CLICKHOUSE_HOST, port=CLICKHOUSE_PORT)
+from vacancies.main.vector_store import (
+    add_vectors,
+    batch_extract_features,
+    embed_features,
+    qdrant_client,
+)
 
 query = """
-SELECT id, chat_username, telegram_id, message, timestamp
+SELECT DISTINCT ON (message) id, chat_username, telegram_id, message, timestamp
 FROM telegram_parser_chatmessage
-WHERE timestamp >= now() - INTERVAL 30 DAY
+WHERE timestamp >= %(timestamp)s
   AND length(message) > 150
   AND arrayCount(x -> position(message, x) > 0, [
     'ваканси', 'ищем', 'требуется', 'разработчик', 'будет плюсом',
@@ -26,7 +30,7 @@ WHERE timestamp >= now() - INTERVAL 30 DAY
   AND arrayCount(x -> position(lower(message), x) > 0, [
     'о себе', 'обо мне', 'умею', '#ищу', '#резюме', 'университет', 'колледж'
   ]) = 0
-  AND id NOT IN %(exist_points)s
+ORDER BY timestamp ASC
 """
 
 
@@ -34,39 +38,23 @@ class Command(BaseCommand):
     help = "Collect vacancies from telegram messages"
 
     def handle(self, *args, **options):
-        next_page_offset = 0
-        exist_points_ids = [-1]
-        while next_page_offset is not None:
-            response = qdrant_client.scroll(
-                collection_name="vacancies",
-                limit=100_000,
-                offset=next_page_offset,
-                with_payload=False,
-                with_vectors=False,
-                timeout=30,
-            )
-            exist_points_ids.extend([point.id for point in response[0]])
-            next_page_offset = response[1]
-        exist_points_set = tuple(set(exist_points_ids))
+        response = qdrant_client.scroll(collection_name="vacancies", limit=1, order_by=OrderBy(key="timestamp", direction="desc"))
+        last_point_timestamp = datetime.now() - timedelta(days=30)
+        if response[0]:
+            last_point_timestamp = response[0][0].payload["timestamp"]
 
-        result_rows = clickhouse_client.query(query, parameters={"exist_points": exist_points_set}).result_rows
-        with ThreadPoolExecutor(max_workers=settings.COLLECT_VACANCIES_BATCH_SIZE) as pool:
-            pool.map(self._process_batch, batched(result_rows, settings.COLLECT_VACANCIES_BATCH_SIZE))
+        clickhouse_client = clickhouse_connect.create_client(host=settings.CLICKHOUSE_HOST, port=settings.CLICKHOUSE_PORT)
+        result_rows = clickhouse_client.query(query, parameters={"timestamp": last_point_timestamp}).result_rows
 
-    def _process_batch(self, result_rows):
-        try:
-            for index, row in enumerate(result_rows):
+        for index, rows in enumerate(batched(result_rows, settings.COLLECT_VACANCIES_BATCH_SIZE)):
+            vacancies_features = batch_extract_features([row[3] for row in rows])
+
+            print(f"Processing {index+1}/{len(result_rows)//settings.COLLECT_VACANCIES_BATCH_SIZE}")
+            with ThreadPoolExecutor() as pool:
+                vacancies_vectors = pool.map(embed_features, [vacancy_features.model_dump() for vacancy_features in vacancies_features])
+
+            for row, vacancy_features, vacancy_vectors in zip(rows, vacancies_features, vacancies_vectors):
                 (id, chat_username, telegram_id, message, timestamp) = row
-
                 link = f"https://t.me/{chat_username}/{telegram_id}"
-                print(f"Processing {index+1}/{len(result_rows)} link: {link}")
-                features = extract_features(message)
-
-                add_vectors(
-                    "vacancies",
-                    id,
-                    features.model_dump(),
-                    {'content': message, 'features_json': features.model_dump(), "link": link, "timestamp": timestamp},
-                )
-        except Exception as exc:
-            traceback.print_exception(exc)
+                payload = {'content': message, 'features_json': vacancy_features.model_dump(), "link": link, "timestamp": timestamp}
+                add_vectors("vacancies", id, vacancy_features.model_dump(), payload, vacancy_vectors)

vacancies/main/vector_store.py

@@ -1,12 +1,9 @@
-from qdrant_client import models
-from langchain_openai import OpenAIEmbeddings
-from langchain_openai import ChatOpenAI
-from qdrant_client import QdrantClient
-from qdrant_client.models import Filter
-from vacancies.main.models import VacancyFeatures
+from langchain_openai import ChatOpenAI, OpenAIEmbeddings
+from qdrant_client import QdrantClient, models
+from qdrant_client.models import Filter, HasIdCondition
+
 from vacancies.conf.settings import QDRANT_URL
-from vacancies.main.models import RecommendedVacancy
-from qdrant_client.models import HasIdCondition
+from vacancies.main.models import RecommendedVacancy, VacancyFeatures
 
 qdrant_client = QdrantClient(url=QDRANT_URL)
 
@@ -16,19 +13,9 @@ FEATURE_NAMES = [
 ]
 
 weights = {
-    "job_title": 25,
-    "employment_type": 5,
-    "work_format": 5,
-    "experience": 8,
-    "position_level": 12,
-    "industry": 10,
-    "tech_stack": 14,
-    "location": 5,
-    "salary_range": 5,
-    "languages": 5,
-    "education": 2,
-    "schedule": 2,
-    "additional_requirements": 2,
+    "job_title": 70,
+    "tech_stack": 10,
+    "salary_range": 10,
 }
 
 vectors_config = {
@@ -38,18 +25,22 @@ vectors_config = {
 if not qdrant_client.collection_exists("vacancies"):
     qdrant_client.create_collection(
         collection_name="vacancies",
-        vectors_config=vectors_config
+        vectors_config=vectors_config,
+    )
+    qdrant_client.create_payload_index(
+        collection_name="vacancies",
+        field_name="timestamp",
+        field_schema="datetime",
     )
 if not qdrant_client.collection_exists("cvs"):
     qdrant_client.create_collection(
         collection_name="cvs",
-        vectors_config=vectors_config
+        vectors_config=vectors_config,
     )
 
 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)
@@ -61,31 +52,21 @@ def _prepare_texts(features):
     return texts
 
 
-def add_vectors(collection_name: str, _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
+def embed_features(features):
+    features = {key: value for key, value in features.items() if value}
+    features_texts = _prepare_texts(features)
+    names, texts = features_texts.keys(), features_texts.values()
+    vectors = dict(zip(names, embedding.embed_documents(texts)))
+    return vectors
 
+
+def add_vectors(collection_name: str, _id: int, features: dict, payload: dict, vectors):
     max_similarities = {}
     for name, vec in vectors.items():
-        if any(v != 0 for v in vec):
-            results = qdrant_client.query_points(
-                collection_name="vacancies",
-                query=vec,
-                using=name,
-                limit=100,
-            )
+        results = qdrant_client.query_points(collection_name="vacancies", query=vec, using=name, limit=100)
         for res in results.points:
-                vid = res.id
-                sim = res.score
-                if vid not in max_similarities:
-                    max_similarities[vid] = {}
-                max_similarities[vid][name] = sim
+            max_similarities.setdefault(res.id, {})
+            max_similarities[res.id][name] = res.score
 
     scored = []
     for vid, feature_sims in max_similarities.items():
@@ -93,77 +74,57 @@ def add_vectors(collection_name: str, _id: int, features: dict, payload: dict):
         scored.append({"id": vid, "score": total})
 
     scored.sort(key=lambda x: x["score"], reverse=True)
-    if scored and scored[0]["score"] > 90:  # threshold
+    if scored and scored[0]["score"] > 80:  # threshold
         return
 
     qdrant_client.upsert(
         collection_name=collection_name,
-        points=[
-            models.PointStruct(
-                id=_id,
-                vector=vectors,
-                payload=payload,
-            )
-        ]
+        points=[models.PointStruct(id=_id, vector=vectors, payload=payload)]
     )
 
 
 def search_similarities(query_filter: Filter, cv_id: int):
-    cv = qdrant_client.retrieve(
-        collection_name="cvs",
-        ids=[cv_id],
-        with_vectors=True,
-    )[0]
+    cv = qdrant_client.retrieve(collection_name="cvs", ids=[cv_id], with_vectors=True)[0]
 
-    max_similarities = {}
-    vacancies_content = {}
+    max_similarities, vacancies_content = {}, {}
     for name, vec in cv.vector.items():
-        if any(v != 0 for v in vec):
         results = qdrant_client.query_points(
             collection_name="vacancies",
             query=vec,
             using=name,
-                limit=100,
+            limit=100000,
             with_payload=True,
             query_filter=query_filter,
         )
         for res in results.points:
-                vid = res.id
-                sim = res.score
-                if vid not in max_similarities:
-                    max_similarities[vid] = {}
-                max_similarities[vid][name] = sim
-                if vid not in vacancies_content:
-                    vacancies_content[vid] = {}
-                    vacancies_content[vid]["content"] = res.payload["content"]
-                    vacancies_content[vid]["link"] = res.payload["link"]
+            max_similarities.setdefault(res.id, {})
+            vacancies_content.setdefault(res.id, {})
+
+            max_similarities[res.id][name] = res.score
+            vacancies_content[res.id]["content"] = res.payload["content"]
+            vacancies_content[res.id]["features_json"] = res.payload["features_json"]
+            vacancies_content[res.id]["link"] = res.payload["link"]
 
     scored = []
     for vid, feature_sims in max_similarities.items():
         total = sum(feature_sims[feature] * weights.get(feature, 1) for feature in feature_sims)
-        scored.append({"id": vid, "score": total, "content": vacancies_content[vid]["content"], "link": vacancies_content[vid]["link"]})
+        scored.append({
+            "id": vid,
+            "score": total,
+            "content": vacancies_content[vid]["content"],
+            "features_json": vacancies_content[vid]["features_json"],
+            "link": vacancies_content[vid]["link"],
+            "sims": feature_sims,
+        })
 
     scored.sort(key=lambda x: x["score"], reverse=True)
 
-    prompt = f"""
-    Резюме: {cv.payload['content']}
-
-    Среди вакансий ниже выбери одну наиболее релевантную и выведи ее индекс(от 0 до 9).
-    Иногда могут попадаться чужие резюме вместо вакансий, их отдавать нельзя.
-    В ответе выведи только число. Если среди вакансий нет подходящих, то верни -1.
-    {scored[:10]}
-    """
-    openai_client = ChatOpenAI(model_name="gpt-5-mini", reasoning_effort="minimal", temperature=0, seed=42, top_p=1)
-    response = openai_client.invoke(prompt)
-    index = int(response.content)
-    if index == -1:
-        return None
-
-    return scored[index]["id"], scored[index]["content"], scored[index]["link"]
+    return scored[0]["id"], scored[0]["content"], scored[0]["link"]
 
 
-def extract_features(content: str) -> VacancyFeatures:
-    prompt = f"""
+def batch_extract_features(contents: list[str]) -> list[VacancyFeatures]:
+    prompts = [
+    f"""
     Extract the following features from the job vacancy description. If a feature is not mentioned, set it to null.
     Features:
     - job_title: Должность (e.g., DevOps, Python программист)
@@ -182,9 +143,11 @@ def extract_features(content: str) -> VacancyFeatures:
     Vacancy content:
     {content}
     """
+    for content in contents
+    ]
     openai_client = ChatOpenAI(model_name="gpt-5-mini", reasoning_effort="minimal", temperature=0, seed=42, top_p=1)
     structured_llm = openai_client.with_structured_output(VacancyFeatures)
-    response = structured_llm.invoke(prompt)
+    response = structured_llm.batch(prompts)
     return response