Project 10 / 10 · RAG & NLP

Text Summarization AI Agent

Autonomous agent on Groq Llama 3.3 70B that detects document domain, picks the summarisation strategy, scores with ROUGE-L and auto-retries on low quality.

01 / Context

The problem

The application treats summarization as a workflow rather than a single prompt, adapting strategy to document domain and checking output quality before accepting the result.

02 / System workflow

From input to output

  1. 01Text / PDF / URL
  2. 02Domain detection
  3. 03Strategy selection
  4. 04Streaming summary
  5. 05ROUGE-L check
  6. 06Retry or export

03 / Architecture

Inside the system

A1

FastAPI accepts text, PDF, URL, batch, comparison, and autonomous-agent requests from a framework-free HTML, CSS, and JavaScript interface.

A2

The agent measures document characteristics, selects a domain-aware prompt and token strategy, calls Groq Llama 3.3 70B, then evaluates ROUGE-L before retrying with an alternate strategy when needed.

A3

PDF ingestion uses pdfplumber, URL extraction uses trafilatura, and token streaming returns the generated summary in real time.

04 / Engineering decisions

What shaped the build

Decision 01

Quality as a loop

ROUGE-L is part of the autonomous path, so a low-scoring first response can trigger another strategy without human intervention.

Decision 02

Multiple work modes

Separate tabs support focused summarization, side-by-side style comparison, batch jobs, browser-side evaluation, and the full agent flow.

Decision 03

Portable delivery

A single Docker image serves the API and interface on the Hugging Face runtime port.

05 / Evidence

Measured, not implied

4

Agent stages

Analyze, select strategy, summarize, and evaluate.

3

Input modes

Plain text, uploaded PDF, or fetched web URL.

5

Workflow tabs

Summarize, compare, batch, evaluate, and agent.

06 / Delivery

What shipped

  • 01The REST surface includes health, model discovery, single, URL, PDF, batch, compare, and agent endpoints.
  • 02Exports cover TXT, JSON, Markdown, and batch CSV; session history remains in the browser.
  • 03GitHub Actions runs tests and linting, then deploys the Docker application to Hugging Face Spaces.

Constraint documented

The autonomous quality gate uses reference-based ROUGE-L, so its usefulness depends on whether reference text is available for the evaluation path.