	In Browser
	StumbleUpon
	del.icio.us
	Google
	Google Buzz
	reddit
	LinkedIn

	Facebook
	Twitter
	Linkedin
	E-Mail

Generative AI > Google Gemini API > Gemini API Text Summarisation at Scale

Gemini API Text Summarisation at Scale

Author: Venkata Sudhakar

Text summarisation is one of the most immediately useful applications of large language models. The Gemini API can condense long supplier contracts, customer feedback threads, support ticket histories, and market research reports into concise summaries - with the key difference from simple extraction being that it understands context and synthesises insights rather than just selecting sentences.

ShopMax India receives lengthy supplier performance reports, legal contract amendments, and customer feedback digests weekly. The operations team used to spend Monday mornings reading through these documents. Now a Gemini summarisation pipeline processes all incoming documents overnight and delivers a structured briefing pack by 8am - turning 4 hours of reading into 20 minutes of reviewing summaries with highlighted action items.

The below example shows how to summarise a long supplier performance report into an executive briefing with action items.

import vertexai
from vertexai.generative_models import GenerativeModel
import json

vertexai.init(project="shopmax-india", location="asia-south1")
model = GenerativeModel("gemini-2.0-flash")

def summarise_document(document_text: str, doc_type: str,
                        max_summary_words: int = 200) -> dict:
    prompt = f"""Summarise this {doc_type} for ShopMax India leadership.
Return ONLY a JSON object with:
- executive_summary: {max_summary_words}-word summary in plain English
- key_findings: list of 5 most important findings or facts
- action_items: list of specific actions required (with owner if mentioned)
- risks: list of risks or concerns identified
- decision_required: true if leadership decision is needed, false otherwise
- urgency: high / medium / low
- one_line: single sentence capturing the most critical point

Document:
{document_text[:12000]}

JSON:
"""

response = model.generate_content(
        prompt,
        generation_config={"temperature": 0.2, "response_mime_type": "application/json"}
    )
    return json.loads(response.text)

# Sample supplier performance report
report = """
SUPPLIER PERFORMANCE REVIEW - Samsung Electronics India - Q1 2024

Delivery Performance: Samsung achieved 87% on-time delivery in Q1, below the
contracted 95% SLA. 23 shipments were delayed by 3-7 days, primarily affecting
Galaxy S24 series stock ahead of the March launch. Two Bangalore store openings
were delayed due to stock unavailability.

Quality: Return rate of 2.1% vs 1.5% target. Main issues: 34 units with display
defects, 18 units with charging port issues. Samsung has acknowledged the display
defect batch and offered credit of Rs 4,20,000.

Pricing: Samsung increased Galaxy S24 Ultra trade price by 3.2% effective April 1
without prior notice, violating the 60-day notice clause in the agreement.

Recommendation: Issue formal notice of SLA breach. Negotiate penalty waiver in
exchange for extended exclusivity on Galaxy S25 series.
"""

summary = summarise_document(report, "supplier performance review")
print(f"ONE LINE: {summary['one_line']}")
print(f"Urgency: {summary['urgency']}, Decision needed: {summary['decision_required']}")
print("Action items:")
for item in summary["action_items"]:
    print(f"  - {item}")

It gives the following output,

ONE LINE: Samsung breached delivery SLA and price notice clause in Q1 2024,
requiring formal notice and renegotiation before April price increase takes effect.
Urgency: high, Decision needed: True
Action items:
  - Issue formal SLA breach notice to Samsung within 5 business days
  - Accept Rs 4,20,000 credit offer for defective display units
  - Negotiate Galaxy S25 exclusivity in exchange for Q1 penalty waiver
  - Escalate price increase violation to legal team for contract review

The below example shows how to process a folder of documents and generate a consolidated briefing pack.

def generate_briefing_pack(documents: list) -> dict:
    briefing = {
        "date": "2024-04-05",
        "urgent_items": [],
        "decision_items": [],
        "summaries": []
    }

for doc in documents:
        summary = summarise_document(doc["content"], doc["type"])
        entry = {
            "document": doc["name"],
            "type": doc["type"],
            "urgency": summary["urgency"],
            "one_line": summary["one_line"],
            "action_count": len(summary["action_items"])
        }
        briefing["summaries"].append(entry)

if summary["urgency"] == "high":
            briefing["urgent_items"].append(doc["name"])
        if summary["decision_required"]:
            briefing["decision_items"].append(doc["name"])

briefing["total_docs"] = len(documents)
    briefing["urgent_count"] = len(briefing["urgent_items"])
    briefing["decision_count"] = len(briefing["decision_items"])
    return briefing

docs = [
    {"name": "Samsung Q1 Review", "type": "supplier performance review",
     "content": report},
    {"name": "Returns Policy Amendment", "type": "legal document",
     "content": "Amendment to returns policy extending window from 7 to 10 days..."}
]

pack = generate_briefing_pack(docs)
print(f"Briefing pack: {pack['total_docs']} docs, {pack['urgent_count']} urgent")
print(f"Decisions needed: {pack['decision_items']}")

It gives the following output,

Briefing pack: 2 docs, 1 urgent
Decisions needed: ['Samsung Q1 Review']

The document text is truncated to 12,000 characters before being passed to Gemini - this handles large documents gracefully while staying well within context limits. For very long documents, a chunking and rolling summary approach works well: summarise each section, then summarise the summaries. ShopMax India processes an average of 15 documents per day through this pipeline, delivering a briefing pack that the operations team reviews in a single 20-minute slot rather than spending the morning buried in reports.

Send your comments, suggestions or queries regarding this site to [email protected].