A Multilingual Research Assistant Is Still an Infrastructure Project

TL;DR for operators

A specialised research platform does not need to discard its documents, metadata, search history, licensing rules, or expert workflows to add an AI assistant. ReSearch_SSH¹ instead proposes a modular layer over the existing ISIDORE infrastructure.

The design combines multilingual domain adaptation with retrieval that connects documents through authors, institutions, themes, citations, and other relationships rather than returning isolated text matches. Most retrieval, reranking, generation, and public knowledge components could be reused elsewhere.

The main portability limit is the component informed by historical queries, clicks, and consultations. Those records reveal which results researchers actually found useful, but they belong to a particular platform and community. Using them may improve alignment with established practice while creating privacy, representativeness, and data-sharing constraints.

The paper reports an architecture and evaluation plan, not effectiveness results. Its corpus is also heavily English-dominant despite covering many languages. Operators should therefore treat the proposal as a blueprint for deciding what to procure, what to build around proprietary assets, what data can be exposed, and what must pass independent evaluation before deployment.

The assistant has to inherit the platform

A specialised research platform rarely begins with an empty database. It already has documents collected under different licences, metadata accumulated through years of curation, search logs, relationships among records, access restrictions, and users with established research practices.

Adding a conversational interface without preserving those assets can make the new system easier to use while making its answers harder to inspect. A fluent response may conceal which documents were retrieved, why they were ranked, which language was used as an intermediary, or whether the model relied on evidence the institution was permitted to process.

ReSearch_SSH starts from a different premise. ISIDORE, the underlying Social Sciences and Humanities discovery infrastructure, remains the documentary foundation. The proposed AI components are layered onto its corpora, metadata, knowledge structures, and historical interactions.

This changes the design problem. The organisation is not merely selecting a capable language model. It is deciding how a model should interact with an existing information system whose records, relationships, legal conditions, and user behaviour already carry institutional knowledge.

Domain adaptation requires several coordinated layers

The paper describes a staged system rather than a single fine-tuning procedure.

First, an open multilingual foundation model would undergo continued pre-training on curated SSH material. The primary resource is an ISTEX subset containing approximately three million documents in more than 60 languages, represented as TEI/XML full text with open metadata. Smaller French, Italian, and bilingual Digital Humanities collections are added to strengthen disciplinary terminology and cross-lingual connections.

Second, retrieval would use more than textual similarity. The system is intended to connect documents through entities and relations such as authors, institutions, topics, citations, and research outputs. Wikidata, OpenAIRE, ISIDORE, and potentially Nakala provide structured context for disambiguation, semantic expansion, navigation, and attribution. This graph-enhanced retrieval and synthesis design is the paper’s use of GraphRAG.

Third, historical searches would help tune ranking. Documents that users clicked or consulted can be promoted, while results that were repeatedly ignored can be demoted. The aim is to align retrieval with observed scholarly practice instead of assuming that semantic resemblance alone captures relevance.

Finally, the generative component would receive instruction tuning for tasks such as multilingual retrieval, comparative document analysis, literature synthesis, and structured state-of-the-art construction.

The resulting architecture can be read as a sequence of distinct alignment problems:

The paper specifies how these layers should interact. It does not yet report which layer contributes most, whether all are necessary, or how their gains compare with a simpler retrieval-augmented system. No completed ablation or benchmark experiment is available.

The behavioural retriever is the least portable component

Much of the proposed stack could, in principle, be reused elsewhere. Another scholarly platform could adopt multilingual domain adaptation, public knowledge-graph enrichment, reranking, source-linked generation, or similar evaluation procedures.

The interaction-aware retriever has a different dependency. It requires historical ISIDORE records linking user queries with clicks, consultations, and ignored results. Those records encode how a particular community has used a particular discovery system.

This makes the component institution-specific in two ways. The data may not exist on a newer platform, and behaviour learned from one scholarly community may not represent another community’s relevance judgements. A history researcher, an economist, and a Digital Humanities scholar can respond differently to the same ranking even when the query terms overlap.

For a platform owner, this produces both an advantage and a constraint. Proprietary interaction records may improve alignment with established workflows and create differentiation that a generic vendor cannot reproduce. They also introduce privacy obligations, sampling bias, retention questions, and portability limits.

Cognaptus inference: organisations assessing specialised research assistants should inventory behavioural data separately from documents and metadata. The decision is not simply whether those records can improve ranking. It is whether they are sufficiently representative, legally usable, and strategically important to justify building a component that cannot be transferred without them.

More than 60 languages does not mean balanced multilingual capability

The corpus description demonstrates breadth, but its token distribution remains highly uneven.

The ISTEX subset contains an estimated 16–21 billion English tokens. French, the largest non-English component, contains approximately 450–600 million. Italian contains only about 10–15 million before the targeted Digital Humanities collections are added.

The additional AIUCD proceedings and Umanistica Digitale materials provide comparatively concentrated Italian and bilingual content. They may improve terminology and connections among Italian, French, and English research traditions. They do not remove the underlying imbalance in the primary corpus.

The initial operational setting is also narrower than the corpus-language count suggests: Italian and English queries over predominantly French scholarly resources. That is a meaningful cross-lingual retrieval problem, but it is not evidence of equivalent performance across the 60-plus represented languages.

A multilingual release decision therefore needs separate tests for language, direction, and task. Italian-to-French retrieval may behave differently from French summarisation, English query expansion, or synthesis across documents written in several languages. Aggregate multilingual scores could obscure those differences.

The source package also preserves a descriptive inconsistency in the paper’s fourth table: the reported English and Italian token rows for Umanistica Digitale sum to approximately two million, while the printed total is one million. This does not alter the architecture, but it reinforces that the corpus figures should be treated as approximate planning data rather than audited training statistics.

Evaluation includes scholarly adequacy, not only model performance

ReSearch_SSH proposes two evaluation tracks.

An independent LLMs4EU team is expected to define quantitative scenarios for retrieval, multi-document summarisation, source traceability, hallucination detection, and cross-lingual performance. French and Italian Digital Humanities experts would then assess scholarly quality, epistemic reliability, methodological adequacy, and usefulness within research workflows.

The expert review is not a decorative user-satisfaction layer. A technically grounded answer can still combine sources in a way that distorts a disciplinary dispute, flattens incompatible methods, or presents an interpretive conclusion with more certainty than the underlying literature permits.

Conversely, expert approval alone would not reveal whether citations reliably support generated claims, whether retrieval fails systematically in one language, or whether the system overlooks relevant documents. The proposed mixed-method design assigns different questions to quantitative testing and disciplinary judgement.

For product and procurement teams, this supports a release gate with at least three distinct decisions:

1. Can the system retrieve and synthesise the relevant material?
2. Can users trace and challenge the resulting claims?
3. Do domain experts consider the output methodologically appropriate for the intended task?

The paper provides the evaluation structure, not the answers. The expert panels are still being formed, and benchmark datasets, model identifiers, annotation procedures, and results remain forthcoming.

Legal controls are part of the technical design

The system would process licensed full text, open metadata, personal interaction records, and generated scholarly synthesis. These inputs carry different permissions and risks.

The paper therefore embeds governance across the development lifecycle through a Data Management Plan, legal and ethics assessment, controlled access, documentation, human oversight, source attribution, sandbox deployment, and case-by-case copyright analysis. Openly accessible material is not assumed to be automatically available for unrestricted model training.

These requirements affect implementation choices. Licensed ISTEX texts may be used within contractual constraints without being released as an unrestricted training dataset. Historical query logs require GDPR-compliant handling. Generated claims need visible links to their sources. Early deployment is planned inside a controlled ISIDORE AI sandbox rather than as an unrestricted public assistant.

This is compliance by design in operational terms: legal permissions, data access, monitoring, attribution, and human challenge mechanisms constrain the system before release. They are not tasks reserved for a final review after training has finished.

What organisations can use before results arrive

The paper directly supports a modular decomposition of specialised research assistance. It identifies separate roles for domain corpora, structured knowledge, behavioural ranking signals, generation, source attribution, expert evaluation, and governance.

Cognaptus infers that this decomposition can improve build-versus-buy decisions. A generic provider may supply the base model and some retrieval components, while the institution retains control over restricted corpora, behavioural data, metadata mappings, evaluation scenarios, and release authority. The relevant procurement unit becomes a stack of replaceable and non-replaceable components rather than one assistant product.

The unresolved issue is effectiveness. The paper reports no completed fine-tuning run, benchmark comparison, expert-panel result, model version, compute requirement, or operational deployment outcome. It also provides no formal evidence that graph grounding or citation removes hallucination.

A reasonable implementation response is therefore a controlled pilot with explicit evidence gates. The design can guide what to build and what to measure. It cannot yet justify claims about retrieval gains, research productivity, reduced error rates, or return on investment.

The blueprint is more mature than the evidence

ReSearch_SSH is valuable because it resists reducing domain AI to corpus size or model choice. It treats specialised assistance as coordination among documents, metadata, user behaviour, structured knowledge, generation, evaluation, and legal authority.

Its sharpest architectural insight is also its principal portability boundary: the retrieval layer closest to actual scholarly behaviour depends on historical records that belong to one infrastructure and community. Its multilingual ambition is similarly bounded by a heavily English-dominant corpus and a narrower initial deployment setting.

For operators, the paper offers a disciplined way to organise the work. Preserve the existing information system. Separate reusable components from institution-specific assets. Test languages and scholarly tasks independently. Require source traceability and expert review. Resolve data permissions before training and deployment.

Whether this produces a better research assistant remains open. The current contribution is a system and governance design precise enough to evaluate, not a validated product ready to copy.

Cognaptus: Automate the Present, Incubate the Future.