Entity Linking: knowing exactly what’s being talked about.
At the Entity Linking challenge of EvalLLM 2026, Kairntech ranked 1st in raw performance — linking every mention in health-surveillance bulletins to the correct official reference record, with a linking F1 of 89.98% (median 75.18%).
Removing ambiguity.
Entity Linking links every mention in a text — a place, a disease, a pathogen… — to the correct record in an official reference database. That’s what lets you know exactly what, or where, is being referred to.
On French health-surveillance bulletins, the mentions were already spotted: the whole difficulty is the liage. Some mentions have no match at all (NIL cases), which is also an expected answer.
Resolution to GeoNames (places) and MeSH (biomedical), a cross-lingual task (FR text, English MeSH) where identifiers can be composite — D001906, D010930:M…, D1 & D2.
Two official references.
The world’s reference geographic database: millions of places with their coordinates, countries and administrative divisions. In case of homonymy, the bulletin’s context decides — the same city name can exist in several countries.
The reference medical thesaurus (diseases, pathogens, toxins…), in its French translation maintained by Inserm, complemented by NLM and NCBI concepts and codes for the finest identifiers.
Three systems, three degrees of control.
The rules allowed three submissions: we used them to compare, on the same task, a continuum from procedural code to autonomous agent.
Gazetteer + LLM
Kairntech’s GeoNames and MeSH indexes propose candidates; a single LLM call decides based on context. The most integrated approach.
Pipeline Python
The code drives everything and calls the LLM only on targeted sub-tasks (translation, paraphrases, final choice). The best trade-off of performance / control / frugality.
Tool-equipped n8n agent
An agent chooses which tools to call and reformulates its queries. Near-complete coverage (best F1), at the cost of much higher inference.
Recall makes the difference.
Our two best systems clearly beat the challenge median. At comparable precision, everything comes down to couverture : the agent leaves barely 2.3% of mentions unlinked.
Kairntech ranks 1st in raw performance; the final ranking then applies frugality criteria that reshuffle the deck — and keep us on the podium.
| System (linking F1) | F1 |
|---|---|
| n8n agent | 89.98% |
| Pipeline Python | 86.69% |
| Gazetteer + LLM | 74.49% |
| Median (19 runs) | 75.18% |
Official results on 200 test documents · 7 teams · 19 runs. 1st in raw performance; final ranking after weighting by frugality criteria.
Precision — and frugality
We measured cost alongside quality. The agent gets the best F1 but consumes nearly twelve times more tokens than the pipeline for just 3.3 points gained. The Python pipeline stays the most frugal and controllable. Frugality — open-source, CO₂ footprint, compute resources — is a criterion we take seriously, and a direction we keep exploring.
Get our Entity Linking resources
Our poster and detailed article from the Entity Linking challenge of EvalLLM 2026: GeoNames & MeSH databases, our 3 systems, results and carbon footprint. Leave your email — we’ll send you everything. (Documents currently available in French only.)
The same precision, working on your data.
Structure and disambiguate your unstructured content with our entity extraction and linking pipelines.