Extract outcomes from narrative reports

Extract outcomes from narrative reports

This processes narrative reports and extracts structured outcome data. Adapt the outcomes framework to match what you track.

from openai import OpenAI
import pandas as pd
import json
from pathlib import Path

client = OpenAI()

# Your outcomes framework - adapt this to your organisation
OUTCOMES_FRAMEWORK = {
    "wellbeing": "Improved wellbeing, mental health, or life satisfaction",
    "employment": "Gained employment, job, or paid work",
    "skills": "Developed new skills, qualifications, or capabilities",
    "housing": "Improved housing situation or accommodation",
    "social_connections": "Reduced isolation or increased social connections",
    "confidence": "Increased confidence, self-esteem, or agency",
    "health": "Improved physical health or access to healthcare"
}

def extract_outcomes_from_report(report_text, report_name):
    """Extract structured outcome data from a narrative report"""

    prompt = f"""You are analyzing a project report to extract outcome data.

OUTCOMES FRAMEWORK (only extract these):
{json.dumps(OUTCOMES_FRAMEWORK, indent=2)}

REPORT TEXT:
{report_text}

Extract all outcomes mentioned that match the framework above. For each outcome found, return:
- outcome_type: which outcome from the framework (use the key name)
- number_of_people: how many people achieved this outcome (extract the number if stated, or "not specified")
- evidence_method: how was this measured? (e.g., "survey", "interviews", "observation", "self-reported")
- quote_from_report: the exact text where this outcome was mentioned
- confidence: your confidence this outcome was achieved (high/medium/low based on evidence quality)

IMPORTANT RULES:
- Only extract outcomes explicitly stated in the report
- Don't infer outcomes not mentioned
- If a number isn't given, note "not specified" not a guess
- Distinguish between "all participants" vs "some" vs a specific number
- If evidence method isn't mentioned, note "not stated"

Return a JSON object with an "outcomes" key containing an array of outcomes found. If no outcomes match the framework, return {"outcomes": []}.

Example format:
{{
  "outcomes": [
    {{
      "outcome_type": "wellbeing",
      "number_of_people": "45",
      "evidence_method": "pre/post survey using WEMWBS",
      "quote_from_report": "45 participants (78%) showed improved wellbeing scores",
      "confidence": "high"
    }}
  ]
}}"""

    response = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{"role": "user", "content": prompt}],
        response_format={"type": "json_object"}
    )

    result = json.loads(response.choices[0].message.content)

    # Add report name to each outcome
    if 'outcomes' in result:
        for outcome in result['outcomes']:
            outcome['report_name'] = report_name

        return result['outcomes']
    else:
        return []

# Process all reports in a folder
reports_folder = Path("project_reports")
all_outcomes = []

print("Extracting outcomes from reports...")

for report_file in reports_folder.glob("*.txt"):
    print(f"\nProcessing: {report_file.name}")

    with open(report_file, 'r', encoding='utf-8') as f:
        report_text = f.read()

    try:
        outcomes = extract_outcomes_from_report(report_text, report_file.stem)

        if outcomes:
            print(f"  Found {len(outcomes)} outcomes")
            all_outcomes.extend(outcomes)
        else:
            print(f"  No outcomes matching framework found")

    except Exception as e:
        print(f"  Error processing: {e}")

# Save to CSV
if all_outcomes:
    df = pd.DataFrame(all_outcomes)
    df.to_csv('extracted_outcomes.csv', index=False)

    print(f"\n{'='*60}")
    print(f"Extraction complete! Found {len(all_outcomes)} outcomes across {len(df['report_name'].unique())} reports")
    print(f"\nOutcome breakdown:")
    print(df['outcome_type'].value_counts())

    print(f"\nConfidence distribution:")
    print(df['confidence'].value_counts())

    print(f"\nSaved to extracted_outcomes.csv")
    print(f"\nNext steps:")
    print("1. Spot-check extractions against original reports for accuracy")
    print("2. Review low-confidence outcomes")
    print("3. Standardize any naming variations")
    print("4. Aggregate numbers (accounting for 'not specified' entries)")
    print("5. Cross-reference with project records for validation")

else:
    print("\nNo outcomes extracted. Check:")
    print("- Are your reports mentioning outcomes from the framework?")
    print("- Is the outcomes framework specific enough?")
    print("- Are report files in the correct folder and format?")