diff --git a/Dockerfile b/Dockerfile
index 28a90d1..d6a2728 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -11,13 +11,14 @@ RUN pip wheel --wheel-dir=/root/.cache/pip/wheels -r /tmp/requirements.txt
 FROM python:3.11-slim
 WORKDIR /app
 
-COPY save_model.py spliting.py requirements.txt ./
+COPY requirements.txt ./
 COPY --from=wheel-builder /root/.cache/pip/wheels /root/.cache/pip/wheels
 RUN pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
 RUN pip install --no-index --find-links=/root/.cache/pip/wheels -r requirements.txt
 
+COPY save_model.py ./
 RUN python3 save_model.py
 
-COPY app.py ./
+COPY app.py spliting.py model.py merge.py ./
 
 CMD uvicorn app:app --host 0.0.0.0 --port 8080 --log-level warning
diff --git a/app.py b/app.py
index 7098954..7d08085 100644
--- a/app.py
+++ b/app.py
@@ -1,32 +1,42 @@
 from spliting import split
 from fastapi import FastAPI
-import torch
-from transformers import pipeline
-
+from model import model
+from spliting import split
+from merge import segmentize_and_merge
 
 app = FastAPI(
     title="Модуль разбиения текстов вакансий",
-    description="Решение представляет собой модуль для разбиения вакансий на должностные обязанности, "
-                "требования к соискателю, условия труда и примечания. Модуль может с легкостью встраиваться "
-                "в информационные системы агрегации с целью повышения качества оформления предлагаемых объявлений."
+    description=(
+        "Решение представляет собой модуль для разбиения вакансий на должностные обязанности, "
+        "требования к соискателю, условия труда и примечания. Модуль может с легкостью встраиваться "
+        "в информационные системы агрегации с целью повышения качества оформления предлагаемых объявлений."
+    )
 )
 
-device = torch.cuda.current_device() if torch.cuda.is_available() and torch.cuda.mem_get_info()[0] >= 2*1024**3 else -1
-model = pipeline("text-classification", "extractor_model", device=device)
 
-
-@app.get("/", description="Принимает на вход текст вакансии, возвращает сгруппированные должностные обязанности, "
-                        "требования к соискателю, условия труда и примечания.")
-async def index(text: str) -> dict[str, str]:
+@app.get(
+    "/",
+    description="Принимает на вход текст вакансии, возвращает сгруппированные должностные обязанности, "
+                "требования к соискателю, условия труда и примечания."
+)
+async def index(text: str, merged: bool = False) -> dict[str, str]:
     result = {"responsibilities": "",
               "requirements": "",
               "terms": "",
               "notes": ""}
+
     text = text.replace("\t", " ").replace("\r", "")
-    sentences = split(text)
-    predicts = [predict["label"] for predict in model.predict(sentences)]
-    for sentence, label in zip(sentences, predicts):
+
+    if merged:
+        segments = segmentize_and_merge(text)
+    else:
+        sentences = split(text)
+        predicts = [predict["label"] for predict in model.predict(sentences)]
+        segments = zip(sentences, predicts)
+
+    for sentence, label in segments:
         result[label.lower()] += sentence + " "
     for key in result.keys():
         result[key] = result[key].strip().replace("  ", " ")
+
     return result
diff --git a/merge.py b/merge.py
new file mode 100644
index 0000000..96935ee
--- /dev/null
+++ b/merge.py
@@ -0,0 +1,95 @@
+from model import model
+from spliting import split
+
+
+def make_segments(text):
+    return [
+        {
+            'text': segment,
+            'scores': {
+                predict['label']: predict['score']
+                for predict in model(segment, top_k=None)
+            }
+        }
+        for segment in split(text)
+    ]
+
+
+def merge_segments(segment1, segment2):
+    segment = segment1['text'] + ' ' + segment2['text']
+    return {
+        'text': segment,
+        'scores': {
+            predict['label']: predict['score']
+            for predict in model(segment, top_k=None)
+        }
+    }
+
+
+def merge_by_sim(segments, bottom_threshold, top_threshold):
+    i = 0
+    while i != len(segments) - 1:
+        conf_1 = max(segments[i]['scores'].values())
+        conf_2 = max(segments[i + 1]['scores'].values())
+
+        try:
+            merged = merge_segments(segments[i], segments[i + 1])
+        except RuntimeError:
+            i += 1
+            continue
+
+        conf_m = max(merged['scores'].values())
+        if (
+            (conf_m > top_threshold) and
+            (conf_1 < bottom_threshold or conf_2 < bottom_threshold)
+        ):
+            segments[i] = merged
+            segments.pop(i + 1)
+        else:
+            i += 1
+
+
+def merge_by_same(segments):
+    i = 0
+    while i != len(segments) - 1:
+        try:
+            merged = merge_segments(segments[i], segments[i + 1])
+        except RuntimeError:
+            i += 1
+            continue
+
+        conf = (
+            max(segments[i]['scores'], key=segments[i]['scores'].get) ==
+            max(segments[i + 1]['scores'], key=segments[i + 1]['scores'].get)
+        )
+        if conf:
+            segments[i] = merged
+            segments.pop(i + 1)
+        else:
+            i += 1
+
+
+def set_low_as(segments, as_, threshold):
+    for i in range(len(segments)):
+        cls = max(segments[i]['scores'], key=segments[i]['scores'].get)
+        if segments[i]['scores'][cls] < threshold:
+            segments[i]['scores'][as_] = threshold + 0.1
+
+
+def segmentize_and_merge(text):
+    segments = make_segments(text)
+
+    merge_by_sim(segments, 0.93, 0.96)
+    merge_by_same(segments)
+
+    set_low_as(segments, 'notes', 0.56)
+
+    result = [
+        [
+            segment['text'],
+            max(segment['scores'], key=segment['scores'].get)
+        ]
+        for segment in segments
+    ]
+
+    return result
diff --git a/model.py b/model.py
new file mode 100644
index 0000000..3a6ef4b
--- /dev/null
+++ b/model.py
@@ -0,0 +1,11 @@
+import torch
+from transformers import pipeline
+
+
+device = (
+    torch.cuda.current_device()
+    if torch.cuda.is_available() and torch.cuda.mem_get_info()[0] >= 2*1024**3
+    else -1
+)
+
+model = pipeline("text-classification", "extractor_model", device=device)