connect ai

ocr/api/message/db_requests.py

@@ -38,4 +38,4 @@ async def save_assistant_user_message(user_message: str, assistant_message: str,
     await settings.DB_CLIENT.messages.insert_one(user_message.to_mongo())
     await settings.DB_CLIENT.messages.insert_one(assistant_message.to_mongo())
 
-    return user_message
+    return assistant_message

ocr/api/message/views.py

@@ -1,7 +1,11 @@
+import asyncio
+
 from ocr.api.message import message_router
 from ocr.api.message.db_requests import get_all_chat_messages_obj, save_assistant_user_message
 from ocr.api.message.models import MessageModel
 from ocr.api.message.schemas import AllMessageWrapper, AllMessageResponse, CreateMessageRequest
+from ocr.api.openai_requests import generate_agent_response
+from ocr.api.report.db_requests import get_report_obj_by_id
 from ocr.api.report.dto import Paging
 from ocr.api.utils import transform_messages_to_openai
 from ocr.core.wrappers import OcrResponseWrapper
@@ -24,9 +28,11 @@ async def create_message(
         reportId: str,
         message_data: CreateMessageRequest,
 ) -> OcrResponseWrapper[MessageModel]:
-    messages = await get_all_chat_messages_obj(reportId)
-    message_history = transform_messages_to_openai(messages)
-    # response = await generate_response()
-    response = 'Hello world'
+    messages, report = await asyncio.gather(
+        get_all_chat_messages_obj(reportId),
+        get_report_obj_by_id(reportId)
+    )
+    message_history = transform_messages_to_openai(messages, message_data.text)
+    response = await generate_agent_response(message_history, report)
     response = await save_assistant_user_message(message_data.text, response, reportId)
     return OcrResponseWrapper(data=response)

ocr/api/openai_requests.py

@@ -0,0 +1,48 @@
+from ocr.api.prompts import OCRPrompts
+from ocr.api.report.model import ReportModel
+from ocr.core.wrappers import openai_wrapper
+
+
+@openai_wrapper()
+async def generate_report(request_content: list[dict]):
+    messages = [
+        {
+            "role": "system",
+            "content": OCRPrompts.generate_report
+        },
+        {
+            "role": "user",
+            "content": request_content
+        }
+    ]
+    return messages
+
+
+@openai_wrapper()
+async def generate_changes(content: list[dict], previous_report: str):
+    messages = [
+        {
+            "role": "system",
+            "content": OCRPrompts.generate_changes
+            .replace("{previous_report}", previous_report)
+        },
+        {
+            "role": "user",
+            "content": content
+        }
+    ]
+    return messages
+
+
+@openai_wrapper()
+async def generate_agent_response(messages: list[dict], report: ReportModel):
+    messages = [
+        {
+            "role": "system",
+            "content": OCRPrompts.generate_agent_response
+            .replace("{reports}", report.report)
+            .replace("{changes}", report.changes or 'There is no changes.')
+        },
+        *messages
+    ]
+    return messages

ocr/api/prompts.py

@@ -0,0 +1,86 @@
+class OCRPrompts:
+    generate_report = """## Task
+
+You must analyze the text extracted from medical document and generate a comprehensive report in **Markdown2** format. Ensure that every detail provided in the document is included, and do not omit or modify any information. Your output must strictly follow the required format.
+
+## Report Structure
+
+The report should be structured as follows, with each section containing only relevant information from the document:
+
+```markdown
+## Patient Information
+
+- Name: [Patient Name]
+- Age: [Patient Age]
+- Date of Scan: [Date]
+- Indication: [Reason for the CT scan]
+
+## Findings
+
+**Primary findings**:
+[Describe significant abnormalities or findings relevant to the indication]
+
+** Secondary findings**:
+[List incidental findings, e.g., "Mild hepatic steatosis noted."]
+**No abnormalities**:
+[Mention organs or systems without abnormalities, e.g., "No evidence of lymphadenopathy or pleural effusion."]
+
+## Impression
+
+[Summarize the findings concisely, e.g., "Findings suggest a primary lung tumor. Biopsy recommended for further evaluation."]
+
+## Recommendations
+
+[Include next steps or further tests, e.g., "PET scan and consultation with oncology recommended."]
+```
+
+[INST]
+
+## Instructions
+
+- **Do not invent or infer any information.** Only use data provided in the user request.
+- Ensure that the format is followed strictly, and the output is complete without any deviations.
+
+[/INST]"""
+    generate_changes = """## Task
+
+You must perform a detailed comparative analysis of the patient's new data from the attached user images against their previous data (`Previous Patient data`). Identify and explicitly highlight all differences, including but not limited to disease progression, remission, newly emerging conditions, and significant clinical changes. Your response must be formatted in **Markdown**.
+
+## Data
+
+**Previous Patient Data**:
+```
+{previous_report}
+```
+
+[INST]
+
+## Mandatory Instructions
+
+- Conduct a **meticulous** comparison of the new and old data, ensuring all discrepancies, updates, and changes in the patient's health status are clearly documented.
+- Provide a structured, concise, and accurate Markdown report.
+- Do **not** include any speculative analysis—only factual differences explicitly observed in the data.
+
+[/INST]"""
+    generate_agent_response = """## Objective  
+
+You are an AI medical assistant. Your task is to provide **precise and direct** answers to the doctor's questions based **only** on the provided `Report`, `Patient changes`, and your **verified medical knowledge**. Your responses must be **brief, factual, and strictly to the point**.  
+
+## Data  
+
+**Report**:  
+```
+{reports}
+```
+
+**Patient changes**:  
+```
+{changes}
+```
+
+## Mandatory Instructions  
+
+- Do not elaborate or provide explanations unless explicitly requested.  
+- **Do not include unnecessary details.** Only provide **essential** information relevant to the doctor's question.  
+- **Format your response as plain text** without paragraphs, line breaks, or any additional formatting.  
+- **Do not speculate.** If the requested information is unavailable in the provided data, respond with: `"Insufficient data to answer."`"""

ocr/api/report/db_requests.py

@@ -21,7 +21,12 @@ async def get_report_obj_by_id(report_id: str) -> ReportModel:
     return ReportModel.from_mongo(report)
 
 
-async def save_report_obj(report: str, changes: str) -> ReportModel:
-    report = ReportModel(report=report, changes=changes, filename='maksim.docx')
+async def save_report_obj(report: str, changes: str | None, filename: str) -> ReportModel:
+    report = ReportModel(report=report, changes=changes, filename=filename)
     await settings.DB_CLIENT.reports.insert_one(report.to_mongo())
     return report
+
+
+async def get_last_report_obj() -> ReportModel | None:
+    report = await settings.DB_CLIENT.reports.find().sort("_id", -1).to_list(length=1)
+    return ReportModel.from_mongo(report[0]) if report else None

ocr/api/report/model.py

@@ -7,7 +7,7 @@ from ocr.core.database import MongoBaseModel
 
 class ReportModel(MongoBaseModel):
     report: str
-    changes: str
+    changes: str | None = None
     filename: str
     datetimeInserted: datetime = Field(default_factory=datetime.now)
     datetimeUpdated: datetime = Field(default_factory=datetime.now)

ocr/api/report/views.py

@@ -1,10 +1,15 @@
+import asyncio
+
 from fastapi import UploadFile, File
 
+from ocr.api.openai_requests import generate_report, generate_changes
 from ocr.api.report import report_router
-from ocr.api.report.db_requests import get_all_reports_obj, delete_all_reports, get_report_obj_by_id, save_report_obj
+from ocr.api.report.db_requests import get_all_reports_obj, delete_all_reports, get_report_obj_by_id, save_report_obj, \
+    get_last_report_obj
 from ocr.api.report.dto import Paging
 from ocr.api.report.model import ReportModel
 from ocr.api.report.schemas import AllReportResponse
+from ocr.api.utils import divide_images, prepare_request_content, clean_response
 from ocr.core.wrappers import OcrResponseWrapper
 
 
@@ -34,8 +39,19 @@ async def get_report(reportId: str) -> OcrResponseWrapper[ReportModel]:
 async def create_report(
         file: UploadFile = File(...),
 ) -> OcrResponseWrapper[ReportModel]:
-    # messages = await create_new_reports(reportId)
-    # response = await generate_response(message_data.text, message_history)
-    report, changes = 'Hello', 'World'
-    report = await save_report_obj(report, changes)
+    try:
+        last_report, contents = await asyncio.gather(get_last_report_obj(), file.read())
+        report, changes = None, None
+        images = divide_images(contents)
+        content = prepare_request_content(images)
+        if last_report:
+            report, changes = await asyncio.gather(
+                generate_report(content),
+                generate_changes(content, last_report.report)
+            )
+        else:
+            report = await generate_report(content)
+        report = await save_report_obj(clean_response(report), clean_response(changes), file.filename)
+    finally:
+        await file.close()
     return OcrResponseWrapper(data=report)

ocr/api/utils.py

@@ -6,10 +6,8 @@ import pytesseract
 from PIL import Image
 from pdf2image import convert_from_bytes
 
-from ocr.api.message.models import MessageModel
 
-
-def transform_messages_to_openai(messages: list[MessageModel]) -> list[dict]:
+def transform_messages_to_openai(messages: list, user_query: str) -> list[dict]:
     openai_messages = []
     for message in messages:
         content = message.text
@@ -17,7 +15,7 @@ def transform_messages_to_openai(messages: list[MessageModel]) -> list[dict]:
             "role": message.author.value,
             "content": content
         })
-
+    openai_messages.append({"role": "user", "content": user_query})
     return openai_messages