The problem at hand

Western democracies produce enormous amounts of data. At Prismos, our goal is to solve the information overload problem that policy experts experience daily by fixing search. We capture and structure an ever-increasing amount of this data, build search algorithms on top, and make it available to our users. One of the ways we make this data available is through a chat interface; that is, an interface where users type instructions rather than click buttons to execute an action. We have been strong believers in the chat interface from the get-go. The very first version of Prismos, which we launched ahead of the Commissioner Hearings in November 2024, already had one, and we have been continuously improving it for about a year now. Users love it because of its simplicity. You just ask questions in plain natural language, and it then becomes our problem to make sure you get the answer you want.

Making it our problem is the only right approach. There should be no learning curve to use Prismos. But take a step back and think about the complexity of this problem. Users could ask literally anything, ranging from very general questions like “which members of parliaments X and Y have been most vocal about topic Z in the past month?”, to very precise questions like “what was said about topic X during meeting Y?”. We can’t simply train an LLM on the data we capture and make that model power the chat interface. LLMs aren’t architected to produce true statements. They predict the most likely word based on the sequence they’ve already generated and the user’s input. If the whole sequence ends up being a true statement, then you’re just lucky. Add to that the humongous cost of having to retrain a model all the time for it to stay up to date with the latest developments in politics, and it quickly becomes clear that we need something else.

What if we treat an LLM for what it is - a smart next-word predictor - and we stop caring about the world knowledge it might or might not store? With that approach, we can split the problem into two: a text generation problem and a data management problem. The LLM takes care of the text generation part. Another piece of software takes care of the data management part. Let’s call the latter one a “query engine”. Essentially, with this approach, the query engine would retrieve those data points from our data infrastructure that contain the answer to the user’s question, and feed those data to the LLM. The LLM should then synthesise these raw data into a coherent whole. Now things become easier to tackle. For starters, we only expect our LLM to excel at synthesising raw data. No more reliance on the knowledge stored in the its neurons. Off-the-shelf LLMs are already good at this, so that means we can save our time and energy and don’t have to bother with training one ourselves. This means we can solve our problem by focusing on the query engine.

Hide and seek

This is where things become interesting. By decomposing our problem in the way we did, we are able to frame it as a Retrieval-Augmented Generation problem (RAG). The RAG paradigm was first introduced in 2020, during the early days of transformer models. RAG applications work like described above: a user asks a question; a query engine retrieves relevant datapoints from a database and feeds them to an LLM; the LLM generates a response based on the context it received; and this output is presented to the user. The very first versions of the Prismos chat assistant were traditional RAG applications along these lines. But we weren’t quite there yet. The main issue, it turned out, was very simple: users could ask literally anything. That means the query engine needs to intelligently pull together data that could, in theory, live in opposite parts of our data infrastructure, often requiring sequential database queries where each query depends on what the previous one returned.

Traditional RAG applications typically retrieve all context upfront in a single pass. They cannot easily handle these complex, multi-step queries. AI engineers often went for the nuclear option to force the query engine to read all information that is possibly relevant, in the hopes that the answer to the user’s query is “somewhere” in that data dump. This doesn’t work quite well. There is only so much context an LLM can handle (see the infamous “context window” limits), and the more context you add, the harder it becomes for an LLM to synthesise it. If you expect someone to find a needle, then don’t hide it in a haystack to begin with. RAG, it turns out, is too rigid. It lacks the flexibility you need when you’re sitting on a lot of data, like Prismos is. We need a just-in-time query engine, software that excels at fetching context exactly when, and only when, this context is needed for the question at hand, and, crucially, excels at forgetting context as soon as it’s no longer relevant to the conversation.

Enter Agentic RAG: an approach to RAG where the query engine is implemented as an agent that dynamically decides which datapoints to read from storage and when. The new Prismos chat assistant is an Agentic RAG assistant. It checks, with every user message, which context needs to be fetched and which context needs to be forgotten. This is why this release is a big deal. You can now ask follow-up questions that require pulling data from multiple sources, or shift topics mid-conversation without the chat assistant getting confused by irrelevant context. Think of mapping stakeholder networks across institutions, tracking how specific issues evolve through different political bodies, or even analysing how a politician’s stance on a topic has shifted over time. All this used to be painstaking manual work, but can now be done conversationally. This isn’t just an incremental improvement. It’s the foundation for an entirely new way of working with political data.

Western democracies are data factories. Governments publish thousands of pages across parliamentary proceedings, regulatory filings, consultation responses, and policy documents. Think tanks and academic institutions release reports and studies. This torrent of information should be democracy's greatest asset; a rich foundation for evidence-based policymaking and informed public debate. Instead, it has become an overwhelming flood that threatens to drown the very insights it contains, and give rise to disinformation instead. Critical connections between disparate pieces of information remain hidden, important developments get lost in the noise, and the policy ecosystem finds itself making choices based on incomplete pictures simply because the relevant information exists but cannot be found efficiently.

The ecosystem has a search problem, and it’s called keyword matching. This crude approach to information retrieval permeates the entire policy sector. Analysts scan documents using CTRL+F. Software tools allow users to input a few key terms, and the system dutifully returns any datapoint that contains those exact words. It's simple, predictable, and works perfectly if you're looking for very specific information; say, any mention of "GDPR" or "carbon tax". But this approach fails spectacularly for almost every other use case. The reason is deceptively simple yet profoundly limiting: synonymy. String matching only matches the very string you specify, and nothing else. If you're tracking discussions about AI but only search for "artificial intelligence", you'll miss conversations about "machine learning" or "neural networks". Each of these terms could be referring to essentially the same policy concerns, but traditional keyword matching treats them as completely unrelated.

Policy and regulatory compliance are notoriously subtle and complex areas where context, nuance, and interconnections matter enormously. Yet the sector relies on search methods that are reminiscent of the early days of Google search, with its Boolean operators, and strip away precisely these qualities. It's an obvious bottleneck that hampers effective governance and informed decision-making, and the main driver why Pepijn and I decided to start Prismos.

Enter semantic search

The "prism" feature is the heart of our software, and we designed it specifically to overcome this fundamental limitation of traditional search. Prisms don't rely on string matching, but on vector search; one of the most remarkable achievements of modern Natural Language Processing. Vectors (sequences of numbers, typically anywhere between a couple hundred to a couple thousand dimensions) can encode meaning in ways that mirror human intuition about language. State-of-the-art deep-learning models will assign nearly identical vector representations to near-synonyms like "artificial intelligence" and "machine learning", while unrelated concepts like "artificial intelligence" and "European Parliament" will be assigned very different vectors.

When you create a prism, you provide a description of what you're looking for - not just keywords, but a richer explanation of the concepts, policies, or issues you want to track. This could be something like "discussions about renewable energy subsidies and their impact on the energy grid" or "debates around data privacy regulations affecting social media platforms". Prism descriptions are turned into vectors, so we’re able to understand what you really mean. We do the same thing with all the political content we process: each meeting fragment, legislative excerpt, policy document, or discussion snippet gets analysed for meaning. When we find content that matches the meaning of what you're looking for, we flag it as relevant to your interests.

This approach solves the synonymy problem elegantly. Your prism about "renewable energy subsidies" will catch discussions about "clean energy tax credits," "solar panel incentives," "wind power support programs," and "green energy funding", even if those exact terms never appeared in your original description. More importantly, it captures the subtle connections and relationships that make policy work so complex. A discussion about "battery parks" might be highly relevant to your renewable energy prism because the underlying concepts are closely related, even if the surface-level keywords seem different.

Instead of playing keyword whack-a-mole, constantly updating your search terms as new phrases emerge, you can describe what you care about in natural language and trust that the system will understand the deeper meaning. Prisms don't solve every challenge in policy monitoring, but they do address the fundamental search problem that has limited traditional approaches. Instead of worrying about whether you're missing information, you can focus on interpreting and responding to the developments and ideas that matter most to your work.

CTRL+F à la poubelle!

Ever since it became clear that SaaS products could form the basis of viable businesses, the majority of such products have essentially been built as opinionated interfaces wrapped around relational databases. These products implemented domain-specific data models (for accounting, marketing, logistics - you name it), made them accessible through an API, and added a polished user interface on top. Companies with a mature data infrastructure (or a very healthy balance sheet) would typically also integrate a machine learning component into the backend stack, often in the form of classifiers acting as adjuncts to the core application logic, making decisions on topics too complex for the engineers to implement manually.

The breakthrough of the type of autoregressive decoders that we nowadays call LLMs didn't change all that much about the way backend systems were built. The generation of LLMs that were first made available to the public in late 2022-early 2023 amazed the world with their ability to generate free-form content, but they were terrible at structured output prediction, or the type of task where the output format actually mattered. And that made it nearly impossible to integrate an LLM into your backend. Even when prompted to generate parseable data (say, JSON or XML), LLMs would often just return broken or malformed structures. I've seen many people initially resort to the "try again and hope it works this time" solution, where backend code would try to parse the output to the requested format and prompt the LLM again if it failed. It was magical thinking with a compute bill - the kind of code that only exists because hype drowned out reason.

The real revolution, in my eyes, came with the commercialization of constrained decoding sometime in Spring 2023 (OpenAI announced theirs in June 2023). LLMs suddenly became capable of producing well-formed, deterministic formats. The impact of this new development cannot be overstated, as it essentially closed the gap between LLMs and traditional ML - not in terms of how they work, but in where they can be used. Suddenly, you could architect software systems where decisions were delegated to a language model with a structured interface, just like you would do with statistical learners.

This is the technology that made a platform like Prismos possible, and this is why we call our product an "AI-first" product. Not because we have an AI agent (we do). It's because we treat LLMs - alongside traditional statistical classifiers - as first-class, legitimate decision-making engines in our core stack. Not an accessory. A foundation.

In the wake of Russia’s full-scale invasion in 2022, the topic of EU enlargement has once again taken centre stage, breaking through more than a decade of ‘enlargement fatigue’ that followed Croatia’s accession in 201.. Fuelled by a renewed sense of geopolitical urgency, not only countries within Russia’s direct orbit (i.e. Ukraine, Moldova and Georgia) but also several Western Balkan countries have made significant progress in their membership trajectories since February 24 2022. In the context of EU enlargement, the European Parliament (EP) plays a crucial yet often underestimated role. Traditionally, the European Commission and the European Council have been the primary actors in this policy domain. However, the EP’s potential to contribute to the debate extends far beyond a simple procedural step in the accession process. The EP’s role in EU enlargement is anchored in both formal powers granted by EU treaties and other (legal) texts, alongside a range of informal powers that have emerged as part of its ‘self-empowerment strategy’ developed over time. In their article "Power beyond borders? Analysing formal and informal tools of the European Parliament in EU enlargement policy," Lien Jansen, Toine Paulissen and Steven Van Hecke (2025) offer a comprehensive examination of five key tools through which the Parliament can leverage its influence over the enlargement process: the consent procedure, budgetary authority, agenda-setting, parliamentary oversight and standing parliamentary delegations.

Among the five identified tools, the consent procedure and the power of the purse stand out as the most prominent. Indeed, one of the European Parliament’s most straightforward tools to weigh on enlargement policy is the consent procedure. Under this legislative procedure, the Council is obliged to consult the Commission and secure the Parliament's consent before moving forward with any membership application. This effectively grants the European Parliament veto power, enabling it to either approve or block a candidate's accession to the EU. However, this veto power is considered to be a rather ‘blunt tool’ as it offers legislation to the EP on a ‘take-it or leave-it basis’, leaving no room to bargain, let alone amend the final text of the accession agreement. Nevertheless, the European Parliament has developed informal practices that help mitigate the formal constraints of the consent procedure. To ensure its concerns are taken into account during the accession negotiations, the EP can, - for instance - adopt resolutions outlining conditions for consent or resort to its power of delay regarding international agreements.

Similarly, the EP’s involvement in the Multiannual Financial Framework (MFF) and the Annual Budget allows it to link funding to specific conditions related to enlargement. Concerning the MFF, the Treaties require the EP to give consent by absolute majority before the Council can adopt the MFF. Additionally, sectoral regulations, which provide the basis for the funding of almost 40 EU spending programs over a seven-year period, are agreed upon under the Ordinary Legislative Procedure putting the EP on equal footing with the Council. This effectively makes the EP a co-legislator on the main external financing instruments regarding enlargement (i.e. IPA and NDICI) as the ceilings and overall amounts of these instruments are decided by the MFF, but not exactly how they are organised, i.e., their general and specific objectives and the size, form and rules of the funding.

It is, however, within the seven-year ceilings of the MFF that the Parliament’s power is most prominent, as it needs to agree with the Council on the Annual Budgets following a special legislative procedure. The Annual Budget is often identified as a very important tool that offers indirect leverage in the area of foreign policy and enlargement and allows the Parliament to steer the political direction of these policy domains. Nevertheless, this power remains at all times constrained by the limited room for manoeuvre left by the MFF. Several scholars (f.e. Corbett et al. 2016; Lord (2017) therefore emphasize that one of the best strategies for the EP to still maximize its (potential for) influence, is aiming at a limited number of priorities in successive annual budgets. However, this strategy is highly dependent on internal cohesion, as conflict seems almost inevitable on a sensitive subject such as the budget, which basically covers all policy areas.

Another key tool concerns the Parliament's agenda-setting power. Through resolutions, (own-initiative) reports, speeches by the EP president and parliamentary debates, the European Parliament can elevate the prominence of EU enlargement, ensuring its remains central to the EU policymaking. At the same time, these instruments enable the EP to bring neglected aspects of the accession process to the forefront. As the public face of the Parliament, the President plays a crucial role in determining whether and how an issue is incorporated into its public messaging. For instance, the current President, Roberta Metsola, frequently underscores the geopolitical necessity of enlargement in her speeches, as illustrated by her recent address at the EU Ambassadors Conference 2025 (analysed using Prismos' chat function). Additionally, the European Parliament plays a crucial role in overseeing the executive, primarily the Commission by scrutinizing enlargement policies through powers as the appointment of the Commissioner for Enlargement (see Prismos’ analysis on the November 2024 hearing of the current enlargement commissioner Marta Kos), debates on statements of the Commission and/or Council, parliamentary questions, information and consultation rights in Association and Accession agreements and, finally, budgetary scrutiny via the discharge procedure. The final key tool is the EP’s network of standing parliamentary delegations to (potential) candidate countries. These delegation serves as an important channel for information exchange and political dialogue, enabling MEPs to gather essential knowledge, which in turn facilitates well-informed decision-making. Additionally, they function as ‘preparatory agents’, laying the groundwork for executive decisions, as they have more space to manoeuvre compared to traditional diplomacy (see for instance Prismos' analysis of the recent EU-Serbia SAPC meeting, particularly in the context of the massive student protests that have erupted across the country)

Despite the EP’s broad range of instruments, the EP’s role in enlargement is not without its limitations. In addition to its relatively limited formal powers, a majority among its members - obviously - appears to be crucial, particularly in the consent and budget procedures. Hence, conflicting interests and political divisions can weaken the Parliament’s bargaining position. Moreover, the overlapping competences between the committees and the standing parliamentary delegations, coupled with the volunteerism in delegation appointments, challenge the effectiveness of these inter-parliamentary bodies. Lastly, the timely and complete information-sharing by the Commission and Council is crucial for the EP’s ability to conduct effective parliamentary oversight in EU enlargement policies (and beyond). Taking these caveats into account, it becomes clear that the European Parliament possesses a range of (in)formal tools to weigh on the accession process. While much depends on the willingness of intra-and inter-institutional political actors to do so, it is an institution that should not be overlooked, especially in the wake of Russia's full-scale invasion and the renewed sense of urgency regarding EU enlargement.

Lien Jansen is a Research Foundation Flanders (FWO) PhD Fellow for Fundamental Research at the KU Leuven Public Governance Institute. Her research focuses on European Politics with a special focus on the European Parliament and EU enlargement towards the Western Balkan countries.

Contact: lien.jansen@kuleuven.be