CineGraph

A movie assistant for factual questions, film recommendations and image retrieval.

Dec 2025 · UZH AI Core project, team of two

Python
RDFLib
SPARQL

Overview

CineGraph is a movie assistant built for the University of Zurich AI Core curriculum. It uses a local RDF knowledge graph to connect films with people, genres and other attributes, so one agent can answer factual questions, recommend films and retrieve images.

Developed with Leyla Khasiyeva.

Interstellar connected to director Christopher Nolan, cast member Anne Hathaway and the science fiction genre. — Illustrative relationships, not a live system result.

What I built

Entity and relation resolution that maps names and wording to graph labels and canonical relations.
SPARQL queries over the local graph, with a TransE embedding fallback when no graph answer is found.
A recommender that scores films with weighted Jaccard over metadata, optionally blended with embedding similarity.
Image retrieval for posters, portraits and scenes, answered through a command line interface or Speakeasy.

CLI / Speakeasy message

agent.py · interactive_agent.py

main.py: CLI without embeddings

`Responser.generate_response()`

Label maps → image intent first → other request routing

Person / company / language rewrites can select the factual branch.

↓ One route per request

Images

Entity + person / film type
IMDb ID → image map
Poster, portrait or scene
Speakeasy image token

Local inputsRDF graph · images.json

Recommendations

Metadata: Weighted Jaccard
Optional embedding cosine similarity
Constraints → ranked film titles
No seeds: theme matching

Local inputsGraph · ratings · entity vectors

Factual questions

Entity + relation → SPARQL
Graph answer → return directly

↓ No answer, and vectors loaded?

TransE fallback

Head + relation → nearest entity

Entity + relation vectors

Local inputsLocal RDF graph + label maps

↓ Route returns its result

Return to CLI / Speakeasy

Formatted answer
Ranked titles
Image token
Explanatory message

Request flow from core/responser.py; the three routes are alternative paths, not consecutive stages.

Results

Three request routes in one agent: factual questions, recommendations and images.
Evaluation checks cover SPARQL prompts with expected answers, plus regex and robustness scripts for media, rate limiting, concurrency and timeouts.
A 6-page evaluation report (8 December 2025) documents the implemented routes and worked examples.

Limits

The checks are focused scripts, not a full agent benchmark.