Skip to content

Agente Auditor ICJCE V2

Agente Auditor ICJCE V2 is the new version of the expert agent on statutory audit and accounting in Spain built for the Instituto de Censores Jurados de Cuentas de España (ICJCE). Like the first version, it answers technical questions in natural language and grounds every answer in its knowledge base: the Spanish Audit Act and its regulation, auditing standards, ICAC rulings, ICJCE guides and circulars, and accounting regulations.

The answer comes structured: the reasoning with the quoted articles, the applicable doctrine, complementary material with checked links and, when the norm allows several readings, the doctrinal disagreement. Compared with the first version, it adds a legal area with case-law search, a coverage plan that makes sure the answer deals with every point of the question, and extended research in trusted internet sources.

When to use it

  • An advisor needs a well-founded answer on auditor independence, with the applicable articles and the ICAC and ICJCE doctrine.
  • A firm wants to know how the courts have ruled on an auditor liability matter.
  • An accounting team asks about the treatment of a transaction under the Spanish General Accounting Plan, ICAC resolutions or IFRS.
  • You want to complete the answer with research in official internet sources on a recent topic.

How to steer the question

The agent decides on its own whether the question is about audit, accounting or both. To force an area, start the message with one of these tags:

TagWhat it does
/AUDIT or /AUDITORIATreats the question as an audit question.
/CONTAB or /CONTABILIDADTreats the question as an accounting question.
/OTRASTreats the question as one about the auditor’s other engagements, other than the audit of financial statements.
/JURIUses the legal area: law applied to audit and case-law search.
/RESEARCH or /RESEARCHFASTPrepares extended research with searches in trusted internet sources.

For example: /JURI ¿Qué responsabilidad tiene el auditor frente a terceros por un informe erróneo?

Configuration

The fields are spread over tabs in the component’s settings panel. These are the ones you need to get started:

FieldWhat it does
Model (required)Connector for the main language model, the one that writes the answers.
Base de datos (required)Connector for the database with the sources. Connect it from a Postgres SQL component with the DatabaseInfo operation.
Embeddings (required)Connector for the embeddings model the database was prepared with, which lets the agent search the sources by meaning.
User PromptThe user’s question. Default: {input}, the message that comes from the chat.
LLM GroundingOptional connector for a fast model that discards irrelevant documents before the analysis.
LLM ComplementarioOptional connector for a fast model that prepares the complementary material. If you don’t connect it, the Model one is used, which may be slower.
LLM ExpertOptional connector for a more capable model that prepares the coverage plan: the points the answer must deal with. If you don’t connect it, no coverage plan is made.
Research LLMConnector for the model that searches the internet in the extended research. It must be a Gemini model; it is needed if the Model one is not.
ToolsOptional connector to give the agent extra tools.
IdiomaLanguage of the answers. Default: es.
Habilitar ContabilidadTurns on the accounting agent for accounting questions. On by default.

Instructions and settings per area

The other tabs hold each area’s instructions and the fine-tuning settings, already prepared for the ICJCE knowledge base:

  • Agentes: the instructions of the audit agent (Auditoría System Prompt) and of the accounting agent, the active specialists and how many source fragments each one uses (the Top K fields).
  • Independencia and Otros Prompts: the instructions for answers on independence, quality and other engagements, and the one for the question classifier.
  • Jurídico: the instructions of the legal area for each type of matter, from E1 — Mercantil to E6 — Organización ICJCE.
  • Research: the format of the research report, the trusted domains (Fast Research — Dominios fiables) and the search and word limits.

Leave them as they are unless the NappAI team tells you otherwise.

Other settings

FieldWhat it does
Agent NameThe agent’s name. Default: Agent.
Agent DescriptionDescription of the agent, useful when another agent uses it as a tool.
Use memoryWhen on, the agent remembers the conversation between messages and understands follow-up questions. Off by default.
StreamWhen on, the answer is shown as it is written. Off by default.

Output

The Response output is the agent’s answer: a structured text with the reasoning, the quoted articles, the doctrine, the complementary material and the sources with their links. If the message is a greeting or is not about its subject, it answers briefly. Connect it to the chat output.

The Agent output delivers the agent itself, to connect it to another component that directs it. When you connect one of the two outputs, the other is hidden.

Example

The ICJCE wants an assistant that answers technical questions with a coverage plan.

  1. Add a chat input component and connect it to User Prompt.
  2. Connect a model component to Model, a more capable one to LLM Expert and an embeddings component to Embeddings.
  3. Add a Postgres SQL component with the DatabaseInfo operation, pointing at the knowledge base, and connect it to Base de datos.
  4. Turn on Use memory and Stream.
  5. Connect Response to the chat output and try “¿Puede un auditor aceptar un encargo de procedimientos acordados en una sociedad en la que su cónyuge es consejero?”.