Preserve event metadata
The ingestion and chunking workflow keeps event identity, dates, location and source references beside text. Retrieval prefers candidates overlapping months mentioned in French or English questions.
Retrieval systems
Case study / 09
Carry source evidence through the retrieval path.
Connect natural-language questions to cultural events through metadata-aware retrieval, grounded answers and source references.
Technologies & concepts
Technical demonstration / selected source
Private repository
01 / The context
Event catalogues contain long descriptions, repeated sessions and inconsistent metadata. Natural-language discovery needs to connect questions to relevant events while preserving dates, locations and the original source links.
02 / Engineering decisions
The ingestion and chunking workflow keeps event identity, dates, location and source references beside text. Retrieval prefers candidates overlapping months mentioned in French or English questions.
A lexical path handles unavailable embeddings or rate limits. When generation is limited, an extractive event list exposes retrieved metadata.
Retrieval recall and reference-based answer judgment are assessed independently. Per-stage instrumentation distinguishes ingestion, cleanup, chunking, embedding, search and generation.
03 / System view
Illustrative architecture from the project presentation. It describes the selected design; it does not imply a deployed service or a measured operational result.
View the full-size diagramEvent catalogue
Normalize and filter
Metadata-aware chunks
Vector retrieval
Answer + sources
04 / Selected artifacts
scripts/fetch_events.pyscripts/preprocess_events.pyscripts/build_vector_store.pyscripts/rag.pyscripts/evaluate_rag.pyscripts/benchmark_pipeline.pydocs/retrieval-design.mddocs/evaluation.md05 / Scope & limitations
Demonstration and source showcase without a hosted assistant, ready-made catalogue, provider credentials, serialized FAISS indexes or historical evaluation results. The retained settings use one example city. Date preference may fill remaining slots outside the requested period and is not a strict filter. Conversation memory and multi-city discovery are outside scope. No accuracy, latency, availability or test-pass claims are made.
The GitHub repository is private. Its link works for authorized collaborators; other visitors can contact me to discuss the project.
Next case study / 10
The next conversation