Podcast Search Engine

A course project for information retrieval that turns a large podcast transcript corpus into a searchable product. The system combines lexical retrieval, semantic vector search, hybrid ranking, and LLM-assisted result exploration in a single web app.

This system was built as part of the project for the course DD2477 Search Engines and Information Retrieval Systems at KTH.

Team Members

• Daniel Ruijs
• Pavlos Spanoudakis
• Bhaswat Raj
• Parniansadat Razaviour

What This Project Does

Podcast transcripts are long, noisy, and difficult to search well with exact keywords alone. This project explores the full retrieval stack:

• lexical search for precision on exact terms,
• vector search for semantic similarity,
• hybrid search to balance precision and recall,
• metadata enrichment to make results easier to scan,
• LLM-powered tools for highlighting, summarization, feedback, and retrieval-augmented question answering.

The result is both a working demo and an evaluation platform for comparing retrieval strategies on the same dataset.

Project Highlights

• Full-stack search experience built with Django, HTMX, and Elasticsearch.
• Three retrieval modes exposed directly in the UI: lexical, vector, and hybrid.
• LLM sidebar for asking follow-up questions over retrieved context.
• Highlight extraction, query-aware summaries, and result-quality feedback.
• Offline evaluation for ranking metrics, highlight quality, and embedding model comparisons.
• Presentation deck included in the repository for a concise project walkthrough.

Demo

Polished UI & layout. 3 retrieval modes provided to the user:

rec1.mp4

Ask a question to the LLM and get an answer using exclusively the results as context:

ask.mp4

Highlight the phrases that are most relevant to the query in each result:

highlight.mp4

Get feedback on the relevancy of each result to the query, as well as overall relevancy metrics:

feedback.mp4

Generate a query-aware summary of the information in the top results:

summarize.mp4

How It Works

The main app lives in engine/ and exposes a transcript search interface with live updates. Users can switch search modes, inspect ranked results, and use the LLM panel to explore the retrieved evidence.

The backend supports:

• query parsing and mode selection,
• search execution across lexical, vector, and hybrid pipelines,
• metadata enrichment before rendering,
• highlight extraction from retrieved snippets,
• summary generation for result sets,
• feedback scoring and RAG-style question answering.

Repository Structure

• engine/ contains the Django backend, HTMX frontend, and LLM-assisted UI.
• indexing/ contains the Elasticsearch indexing scripts and embedding benchmarks.
• evaluation/ contains ranking, metric, and highlight evaluation scripts.
• embedding eval/ contains embedding-model evaluation experiments and generated outputs.

Quick Start

1. Start Elasticsearch

Install Docker, then run:

curl -fsSL https://elastic.co/start-local | sh

This creates a local Elasticsearch and Kibana setup and an elastic-start-local directory with the credentials needed for the rest of the project.

After installation:

• Elasticsearch: http://localhost:9200
• Kibana: http://localhost:5601

Test the connection with:

cd elastic-start-local/
source .env
curl $ES_LOCAL_URL -H "Authorization: ApiKey ${ES_LOCAL_API_KEY}"

2. Run the Web App

From the engine directory:

cd engine
uv sync
uv run python manage.py check
uv run python manage.py runserver

Open the app at:

http://127.0.0.1:8000/

3. Configure Required Environment Variables

Create engine/web/services/.env with:

API_KEY=your_elasticsearch_api_key
METADATA_TSV_PATH=/absolute/path/to/metadata.tsv
GEMINI_API_KEY=your_gemini_api_key
GEMINI_MODEL=gemini-2.5-flash-lite
GEMINI_FEEDBACK_MODEL=gemini-3.1-flash-lite-preview

4. Index the Dataset

Follow the instructions in indexing/README.md to load the podcast transcript corpus into Elasticsearch before using the demo end to end.

Evaluation and Experiments

This repository includes reproducible evaluation workflows for the retrieval stack:

• ranking evaluation across lexical, vector, and hybrid retrieval,
• highlight-quality evaluation for LLM-generated snippets,
• embedding benchmark scripts,
• LLM-judged embedding retrieval experiments,
• saved metrics and result files for comparison.

Good entry points:

• indexing/README.md
• evaluation/README.md
• embedding eval/README.md

Local Elasticsearch Setup

If you want to manage the local Elasticsearch instance manually:

cd elastic-start-local
./start.sh

To stop it:

cd elastic-start-local
./stop.sh

To uninstall it completely:

cd elastic-start-local
./uninstall.sh