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

	Facebook
	Twitter
	Linkedin
	E-Mail

Generative AI > RAG Pipelines > AI Document Summariser

AI Document Summariser

Author: Venkata Sudhakar

A typical vendor contract, legal agreement, or business proposal runs 30-50 pages. An executive needs a 1-page brief. A procurement manager needs the key commercial terms. A lawyer needs the liability clauses highlighted. Reading the full document each time is not practical when you receive dozens per week. An AI document summariser reads the entire document and produces a structured, role-specific summary in seconds - the same document can produce a 5-bullet executive brief, a detailed clause-by-clause breakdown for legal, and a commercial terms sheet for finance, all from one pipeline.

Long documents exceed the LLM context window, so a direct summarise-everything approach fails for documents over 20-30 pages. The map-reduce pattern solves this: first map each chunk of the document to a short chunk summary, then reduce all chunk summaries into a final cohesive brief. LangChain provides this as a built-in chain (load_summarize_chain with chain_type="map_reduce"), but you can also build it explicitly with LCEL for full control over the prompts at each stage - giving you different summaries for different audiences from the same source document.

The below example summarises a long vendor software contract using the map-reduce pattern, then produces three role-specific views: an executive brief, a commercial terms sheet, and a risk flag list for legal.

from langchain_openai import ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser
from langchain.text_splitter import RecursiveCharacterTextSplitter

llm = ChatOpenAI(model="gpt-4o-mini", api_key="your-api-key", temperature=0)

# Simulated vendor contract text (in production, load from PDF or DOCX)
contract_text = """
SOFTWARE SERVICES AGREEMENT between TechVendor Solutions Pvt Ltd (Vendor)
and GlobalRetail India Ltd (Customer), dated 1 April 2025.

CLAUSE 1 - SERVICES: Vendor agrees to provide cloud-based inventory management
software including implementation, training, and 24x7 support for 3 years.
Initial implementation is estimated at 90 days from contract signing.

CLAUSE 2 - FEES: Annual license fee of Rs 24,00,000 payable quarterly in advance.
Implementation fee of Rs 5,00,000 payable 50 percent on signing and 50 percent
on go-live. Support charges of Rs 2,40,000 per annum included in license fee
for first year. Price escalation capped at 8 percent per year from year 2.

CLAUSE 3 - SLA: System uptime guaranteed at 99.5 percent monthly excluding
scheduled maintenance. Downtime credits of 10 percent monthly fee per hour
below 99.5 percent, capped at 30 percent of monthly fee.

CLAUSE 4 - DATA: Customer retains full ownership of all data. Vendor may not
use Customer data for any purpose other than service delivery. Data export
available in CSV and JSON formats within 30 days of contract termination.

CLAUSE 5 - LIABILITY: Vendor total liability capped at 6 months of fees paid.
No liability for indirect or consequential loss. Force majeure clause applies
to both parties for events beyond reasonable control lasting over 30 days.

CLAUSE 6 - TERMINATION: Either party may terminate with 90 days written notice
after Year 1. Customer may terminate immediately for material breach uncured
within 30 days of written notice. Early termination fee of 3 months fees
applies if Customer terminates without cause in Year 1.

CLAUSE 7 - INTELLECTUAL PROPERTY: All custom development paid for by Customer
becomes Customer property. Base platform and all underlying IP remains Vendor
property. Customer granted perpetual licence to use custom modules post-termination.

CLAUSE 8 - GOVERNING LAW: This agreement is governed by laws of India.
Disputes subject to arbitration in Bangalore under Indian Arbitration Act.
"""

splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=100)
chunks = splitter.split_text(contract_text)
print(f"Document split into {len(chunks)} chunks")

Map-reduce summarisation with three role-specific outputs,

# STEP 1: MAP - summarise each chunk
map_prompt = ChatPromptTemplate.from_template(
    "Summarise the key points from this contract section in 2-3 bullet points:\n\n{chunk}"
)
map_chain = map_prompt | llm | StrOutputParser()

chunk_summaries = [map_chain.invoke({"chunk": c}) for c in chunks]
combined = "\n\n".join(chunk_summaries)

# STEP 2: REDUCE into three role-specific views from the same combined summary

exec_prompt = ChatPromptTemplate.from_template(
    "From these contract notes, write a 5-bullet EXECUTIVE BRIEF covering:\n"
    "what is being purchased, total cost, contract duration, key risk, key benefit.\n\n{notes}"
)

commercial_prompt = ChatPromptTemplate.from_template(
    "From these contract notes, create a COMMERCIAL TERMS TABLE with:\n"
    "Annual fee, Implementation fee, Payment terms, Price escalation, SLA, "
    "Termination notice, Early exit penalty.\n\n{notes}"
)

legal_prompt = ChatPromptTemplate.from_template(
    "From these contract notes, list the TOP LEGAL RISK FLAGS as numbered points.\n"
    "Focus on: liability caps, indemnity gaps, termination risks, IP issues, "
    "data concerns.\n\n{notes}"
)

reduce_chain_exec       = exec_prompt       | llm | StrOutputParser()
reduce_chain_commercial = commercial_prompt | llm | StrOutputParser()
reduce_chain_legal      = legal_prompt      | llm | StrOutputParser()

exec_brief   = reduce_chain_exec.invoke({"notes": combined})
comm_terms   = reduce_chain_commercial.invoke({"notes": combined})
legal_flags  = reduce_chain_legal.invoke({"notes": combined})

print("=== EXECUTIVE BRIEF ===")
print(exec_brief)
print("\n=== COMMERCIAL TERMS ===")
print(comm_terms)
print("\n=== LEGAL RISK FLAGS ===")
print(legal_flags)

It gives the following output,

=== EXECUTIVE BRIEF ===
- PURCHASE: 3-year cloud inventory management software from TechVendor Solutions
- COST: Rs 24L/year license + Rs 5L one-time implementation = ~Rs 77L over 3 years
- DURATION: 3 years with 90-day termination notice permitted after Year 1
- KEY RISK: Vendor liability capped at 6 months fees - limited recourse for major failures
- KEY BENEFIT: Full data ownership, perpetual licence for custom modules, uptime SLA

=== COMMERCIAL TERMS ===
Annual Fee:          Rs 24,00,000 (paid quarterly in advance)
Implementation Fee:  Rs 5,00,000 (50% on signing, 50% on go-live)
Support:             Included in Year 1; Rs 2,40,000 p.a. from Year 2
Price Escalation:    Capped at 8% per year from Year 2
SLA Uptime:          99.5% monthly; credits up to 30% of monthly fee
Termination Notice:  90 days written notice after Year 1
Early Exit Penalty:  3 months fees if terminated without cause in Year 1

=== LEGAL RISK FLAGS ===
1. LIABILITY CAP: Vendor liability limited to 6 months of fees paid - inadequate
   for a critical system failure causing major business loss.
2. NO CONSEQUENTIAL LOSS: No recovery for lost profits, reputation damage,
   or business disruption even if caused by Vendor negligence.
3. FORCE MAJEURE: 30-day threshold is generous to Vendor - could block SLA credits
   during extended outages framed as force majeure events.
4. TERMINATION CURE WINDOW: 30 days to cure material breach may be too long
   if breach involves a data security incident.

The same pipeline works for any long business document: board meeting minutes, tender documents, annual reports, HR policies, or insurance certificates. The map step is fast and cheap because chunk summaries are short. The reduce step is where your business-specific prompt makes the difference - the more specific your questions about what to extract, the more useful the output. For a procurement team handling 50 contracts per month, this pipeline replaces 200+ hours of reading time with a 10-minute automated review.

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