Proposal · RFP v2.1
8. Search and AI functionality
Chrysal Website 2027 · Extended draft · 19 October 2026
RFP-hoofdstuk 08
Search before AI for AI’s sake
Good conventional search functionality at launch. AI only when a validated user need justifies it.
Our response & approach
Searching with relevant filters
At launch, we recommend a good conventional search function with relevant sorting, synonyms and filters. Visitors must be able to find products, knowledge, cases and technical documentation, with recognisable labels and attention to language and market. We test search quality with fifty representative search queries, including technical terms, product names and alternative spellings.
Managed knowledge base as a foundation
At the same time, the platform is set up as a ‘second brain’: a managed collection of approved knowledge with sources, relationships, versions and content owners. We record texts and document metadata in a structured way and enable a documented export or API access. This is the preparation for later AI access. Adding content does not in itself train a language model; updating a knowledge index and any model training are different processes.
Assessing the current assistant Lily
Lily is already visible as a digital assistant on the current public website. During discovery, we investigate what content Lily uses, who owns the technology, what usage data are available and what works well or less well. The choice is then to retain it with a demonstrably secure connection, temporarily pause it or replace it later. Taking over or integrating Lily is not an implicitly included licence or technical connection. [2]
RAG as a possible next step
For a subsequent phase, we recommend, if necessary, an AI assistant that first retrieves approved knowledge and only then formulates an answer. This approach is called retrieval augmented generation, or RAG. Answers display source citations, take account of market and document version and refer to a human if the sources are insufficient. For product dosage, safety information and application, we do not allow unsubstantiated advice to be generated. Chrysal expertise remains recognisable as the source.
Secure configuration and bounded answers
The AI layer is connected through a protected integration, with server-side keys, usage limits, privacy agreements and a test set. An example: someone asks about an application outside the approved documentation. The assistant states that it has no validated answer and displays the local expert, rather than devising application advice itself. A pilot only goes live after substantive review and checks for unwanted instructions, data leaks and incorrect answers.
Search and knowledge base at launch
At launch, we deliver search, the structured knowledge base and the documentation for a later connection. A new AI assistant, vector index, model usage and ongoing answer monitoring are explicit options in chapter 16. This first establishes the content quality on which AI can truly rely.
Save this page as PDF
The print version is formatted as a document, without buttons or navigation. Choose “Save as PDF” in the print dialog.