The words we use,
defined from production.

The European regulation on artificial intelligence, imposing obligations proportionate to the risk of the use: documentation, decision traceability, bias and performance checks.

The pitfall These are built in at design time. Retrofitting them onto a deployed system costs more than planning for them.

A system that calls tools and changes the state of another system: it creates an entry, sends a message, updates a record. The difference from a conversational assistant is not conversation quality, it is consequence.

The pitfall An agent is judged on its rate of correct actions and the cost of a wrong one, not on the relevance of an answer.

The billing unit of agents: each successful action by an agent (a question resolved, a document processed, a record prepared) uses one credit. The vendor sets the credit price and bills it to its customers, under its brand; the platform provides the per-customer meter and packs.

The pitfall The customer only pays for work done, not one more licence. Common rule of thumb: bill about 10% of the value of the work saved, the customer keeps the rest of the gain.

The description of entities, their relations, the states a record can take and the permitted transitions between them, with the mandatory checks at each step.

The pitfall It replaces "the model will figure it out" with "here is what is permitted". A free agent finds a correct path on standard cases and invents one on rare cases.

Cutting a document into segments before embedding them. Each segment must be understandable on its own.

The pitfall Cutting every n characters is the number one cause of incomplete answers: it separates a table header from its rows, a condition from its exception. Chunk along the structure instead.

The value below which the system does not act alone and hands over to a human.

The pitfall It is set on the cost of an error in each direction, not on an accuracy figure: missing a good case and accepting a bad one almost never carry the same price.

How much text a model can take into account at once, measured in tokens.

The pitfall A large window does not replace retrieval: filling the context with everything you have degrades accuracy and multiplies cost. What matters is selecting, not piling up.

The representation of a text as a numeric vector, such that two texts with close meanings produce two close vectors. It is what makes similarity search possible.

The pitfall Embeddings are excellent on meaning and poor on identifiers: a product reference or a proper noun has no reliable semantic neighbourhood.

A set of real cases with their expected result, produced by business experts, used to measure whether a change to the system improves or degrades it.

The pitfall Fifty to two hundred cases are enough, but they must include the ambiguous ones and come from users. Without an evaluation set, every change is a bet.

Retraining an existing model on your data to change its behaviour: an output format, a tone, a response structure.

The pitfall It is not an alternative to RAG. RAG solves a knowledge problem, fine-tuning a behaviour problem. A need for up-to-date knowledge is not addressed by retraining.

An approach where knowledge is structured as a graph rather than as text chunks, so a question becomes a traversal of relations instead of a similarity search.

The pitfall Useful when the answer exists in no single paragraph and must be reconstructed by following a chain, why an invoice is blocked, for instance.

The controls placed around a model to bound what it can receive and produce: input and output schema validation, filters, tool perimeter, explicit refusal out of scope.

An answer produced by a model that is plausible and wrong. The term is misleading: nothing malfunctioned, the model did exactly what it was asked, produce the most likely continuation.

The pitfall The remedy is not a better model, it is an exit: a system that always answers will always produce wrong answers without flagging them.

A platform that provides the whole engine through APIs without imposing an interface: the vendor keeps the "head", meaning the screen, the brand and the relationship with its users, and plugs the agents underneath.

The pitfall It is not a white-label agent sitting next to the software: the agent works inside the vendor's screen, under its name, and the user never sees Rakam.

The arrangement by which a system's proposal goes in front of a person before being applied, with its justification and its score.

The pitfall An irreversible action must go through it whatever the score. That is the rule that makes an agent deployable.

The combination of vector search and lexical keyword search, with the two rankings fused.

The pitfall It is almost always the change that buys the most for the least engineering, because the two methods fail on different cases.

The practices for operating a language-model system in production: continuous evaluation, traces, latency and cost tracking, drift detection, prompt version management.

A protocol standardising how a model discovers and calls external tools: tool descriptions, the shape of the call, and the exposure of read-only resources.

The pitfall It is not a layer of intelligence, it is a standardised socket. It does not decide when to call what, that decision stays with the agent.

A model whose parameters are published and can be run on your own infrastructure, including air-gapped networks.

The pitfall It is what makes sovereignty practical, and what puts a ceiling on the price of closed models.

Counting usage customer by customer inside a shared system.

The pitfall It is the mechanism that lets a software vendor bill an AI feature on usage. Without it, the feature stays demonstrable but not sellable.

A demonstration meant to verify that an approach is feasible, without the requirements of a production system.

The pitfall A POC that never ships usually has three causes: no acceptance criteria, no usage metering, no revenue target. All three are set at the start, not afterwards.

An attack where content read by the model, a ticket, a comment field, a CV, a forwarded email, contains instructions it treats as directives.

The pitfall No prompt tuning solves this reliably. The protection is structural: closed tool perimeter, human validation on the irreversible, content marked as content.

At question time, the relevant pieces are retrieved from your data and placed in the model's context, so it answers from those pieces rather than from what it memorised.

The pitfall The point is not only accuracy, it is traceability: an answer built from identified chunks can cite its sources.

Reordering the retrieved segments with a finer model that judges actual relevance rather than mere vector proximity.

The parameter controlling how much randomness enters generation. At zero, the model produces the most likely continuation almost deterministically.

The pitfall On a business system, a high temperature is almost always a mistake: two runs of the same question must give the same result.

The unit of text segmentation a model works with, smaller than a word. It is the billing unit of most providers.

The mechanism by which a model requests the execution of a declared function, with parameters it produces that are validated before execution.

The pitfall The quality of the tool descriptions matters more than the choice of model. A badly named tool will be called at the wrong moment.

A store that indexes vectors so the closest ones to a given vector can be retrieved quickly, along with their metadata.

The pitfall Below a few hundred thousand chunks, a vector extension on the database you already run is plenty and saves you a component to operate.

An agent that does the work inside business software: it reads the state of a record, calls the APIs under existing permissions, follows a path approved in advance, and hands over to a human where the cost of a mistake warrants it.

The pitfall It is not one more agent type in a product line. Support, SalesOps, SmartData and AdminOps are the same engine with a different graph and different tools, which is why one can be configured that appears on no list.

An agent's operational brain: the steps of a business task, in order, with what it checks at each step, the tools it may call and where a human takes over. The agent follows this path instead of improvising it.

The pitfall It is what makes an agent predictable and testable. Our graph approach comes from our research, published at ICMLC 2026.