Research
Relevant is not always valid.
A retrieval system can return evidence that matches a question and still be wrong about when it was true: a medication that was stopped, a symptom that resolved, a procedure that was only planned. My research asks how retrieval can account for when evidence holds, not only what it is about.
How this came together
- Sept. 2025
Télécom SudParis: Research internship on temporal retrieval-augmented generation for healthcare.
- Undated
Master's thesis: A structured state-of-the-art review of temporal retrieval-augmented generation, covering methods, evaluation, and open challenges. The temporal-validity research gap was identified while writing it.
- 22 June 2026
State Engineer thesis defended: “Building a Temporal Retrieval-Augmented Generation System for Healthcare Applications”.
- June 2026 · in development
Developing TempMedBench, a synthetic patient-timeline retrieval benchmark and generation pipeline.
Alongside the thesis: the temporal annotation pipeline. Its annotation output contributed data used for the work that became WaTE.
Current and selected work
Measuring it
Benchmark · in development · June 2026
TempMedBench
A synthetic patient-timeline retrieval benchmark and generation pipeline for temporally faithful healthcare retrieval and reasoning. Structured longitudinal patient timelines are built from synthetic patient records, then used to generate temporally dependent questions, gold evidence relevance judgements, and model-ready note chunks against which retrieval strategies are evaluated.
Role
The benchmark idea and core contribution are Chamel's. TempMedBench is separate from WaTE and from the temporal annotation pipeline; it grew out of the broader Temporal RAG thesis work.
Retrieving under temporal constraints
Thesis research · Télécom SudParis · Sept. 2025 – June 2026
Temporal RAG for Healthcare
A retrieval-first architecture that combines semantic relevance with explicit temporal constraints over longitudinal healthcare information, addressing temporal granularity, event-time anchoring, provenance, and retrieval-time consistency.
Fig. 2 Excerpt of the patient-scoped retrieval pipeline: semantic and temporal scores are combined to rank evidence. Simplified from the State Engineer thesis; the full figure is in the case study. No scores. The data underneath
Research software · Télécom SudParis · Sept. 2025 – June 2026
Temporal annotation pipeline
A temporal annotation pipeline extended from temporal-expression detection and normalization to event extraction, event-to-time linking, and document-creation-time anchoring. The annotator combines rules, spaCy, and a small language model.
Role
Held primary responsibility for the pipeline. The TIMEX stage existed previously in Java; migrated it to Python and extended the pipeline. Independently proposed the synthetic temporal-document direction and built the generation pipeline.
Representing time
Paper · 2026
WaTE: Continuous Distributional Temporal Representations for Text Encoders
Role
Co-author. Primary responsibility for temporal annotation and data construction; contributed tokenizer-related ideas, benchmark selection (TempEval-3, TimeBank-Dense, TS-RETRIEVER), evaluation methodology, ablation studies and controls, and manuscript preparation. Training strategy, tokenizer design, and ablation methodology were joint research decisions; the final pretraining experiments and reported results were carried out by collaborators.
Papers & theses
Paper · 2026
WaTE: Continuous Distributional Temporal Representations for Text Encoders
Accepted short paper, 38th IEEE International Conference on Tools with Artificial Intelligence (ICTAI 2026).
State Engineer thesis · 2026
Building a Temporal Retrieval-Augmented Generation System for Healthcare Applications
ESI Algiers and Télécom SudParis, defended 22 June 2026.
Master's thesis · 2026
Exploring Temporal Retrieval-Augmented Generation: Methods, Evaluation, and Open Challenges
ESI Algiers, defended 22 June 2026.
Open questions
If a retriever does well on time-dependent questions, how much of that success rests on evidence that is relevant but invalid for the time the question asks about, and would an aggregate score ever show it?
When a retriever gets a question about a particular time right, how can we tell whether it accounted for when the evidence holds, rather than relying on a shortcut such as matching an explicit date?
Which temporal situations, such as a state that has ended or an event that was only planned, are poorly covered by the data we use to build and evaluate retrieval systems, and which failures on those cases does an average score conceal?
How should a retrieval system reason about a fact whose validity interval is only partly known, for example when its start is recorded but its end is implied only by a later record?