How do Ecodex’s AI agents work?
We can separate Ecodex’s 4 AI agents into 2 categories:
- Documentation search agents, namely the Doc agent and the Methodo agent: these agents, based on a question asked by the user, will be able to extract relevant answer elements (text, formulas, tables, etc.) scattered within certain documents (generally .pdf files), and then present them intelligibly to the user.
- Database search agents, namely the FE agent and the Decarbo agent: these agents will *transform the user’s question into a query for the search engine, then analyze and present the answers returned by the engine as synthesized answer elements.
Examples:
. Interaction with the FE agent, the user asks their question: « What is the carbon impact of bicycle raw materials? »
. The agent converts the request into keywords for the search: « steel, aluminum, etc. »
. The search engine returns a set of answers that match the keywords, in descending order of relevance: (1) « New steel or tinplate, Carbon Base », (2) « Recycled steel or tinplate, Carbon Base », etc.
. The agent takes over one last time, synthesizes the different proposals from FE and offers a structured answer: « I suggest you use such FE from the Carbon Base, etc. »
In technical terms, this is known as RAG (Retrieval-Augmented Generation): this approach solidifies the response quality of a generative AI, which therefore has much lower probabilities of hallucinating, as it is directly « connected » (augmented) to a reference source (i.e., public reference carbon documentation or databases).
To further enhance the quality of the proposed answers, the language models used by the Doc and Methodo agents do not search directly in the raw files - which would happen, for example, if you asked a chatbot to navigate directly within documentation, and generally leads to limited quality answers.
Several preliminary steps of digitization (OCR), parsing, and extraction have been carried out for each document to transform them into structured elements manageable by the language models.