Skip to content

20.md

Latest commit

 

History

History
134 lines (97 loc) · 3.54 KB

20.md

File metadata and controls

134 lines (97 loc) · 3.54 KB

Prescription Drug Ad Review: Scene Implementation

Date: 2024-12-05 Topic: Implementing a regulatory compliance scene with VL+LLM

Background

Today I implemented a scene for reviewing prescription drug advertisements - checking if they contain required compliance statements.

The Compliance Requirement

Prescription drug ads must prominently display: "This advertisement is for medical and pharmaceutical professionals only."

The scene needs to:

  1. Extract text from the advertisement image/PDF
  2. Check if the required statement is present and prominent
  3. Return a compliance judgment

Simplified Implementation

The initial implementation was overly complex. After refactoring:

@router.post("/prescription_ad_review", response_model=BaseJudgementResponse)
async def recognize_prescription_ad_api(
    advertisement_file: UploadFile = File(...),
    batch_size: int = Form(default=1)
) -> BaseJudgementResponse:
    vl_prompt = """Please identify if there is a prominent statement
    'This advertisement is for medical and pharmaceutical professionals only'..."""

    try:
        processed_doc = await recognizeImageTextVL(
            advertisement_file,
            prompt=vl_prompt,
            batch_size=batch_size
        )
        content = "\n".join([page["content"] for page in processed_doc["pages"]])
        result = await llm_chat(query=analysis_query)

        return BaseJudgementResponse(
            code=200,
            msg="Recognition successful",
            data=BaseJudgement(
                is_illegal="not prominently stated" in result,
                reason=result
            )
        )
    except Exception as e:
        return BaseJudgementResponse(
            code=400,
            msg=f"Processing failed: {str(e)}",
            data=None
        )

Supported File Types

SUPPORTED_IMAGE_TYPES = frozenset({
    'image/jpeg',
    'image/png',
    'image/gif',
    'image/bmp',
    'application/pdf'
})

The batch_size parameter controls PDF page processing - balance between speed and resource usage.

Error Handling

Standardized error responses with exception handlers:

@app.exception_handler(RequestValidationError)
async def validation_exception_handler(request, exc):
    return JSONResponse(
        status_code=422,
        content=BaseResponse(
            code=422,
            msg="Validation error",
            data={"detail": exc.errors()}
        ).dict()
    )

SSH Connection Troubleshooting

Also spent time debugging SSH access to the deployment server:

# Check service status
systemctl status sshd

# Test with verbose output
ssh -vvv username@hostname

# Verify firewall
sudo iptables -L

The issue was a firewall rule blocking the SSH port.

Today's Reflection

The refactoring process was satisfying. The original code had multiple layers of unnecessary JSON parsing and data model conversions. The simplified version does the same thing in half the code.

The key insight: start simple, add complexity only when needed. The original implementation anticipated edge cases that never occurred, adding complexity without value.

The SSH debugging was a reminder that deployment issues are often infrastructure, not code. Knowing basic network troubleshooting (firewall rules, service status, connection testing) is essential for full-stack development.

Further Learning

  • Regulatory text analysis with NLP
  • Image text detection (OCR) improvements
  • FastAPI middleware patterns
  • Infrastructure automation with Ansible