The price of knowledge base is first of all a matter of scope
Let’s start with the conclusion: there is no fixed price for enterprise AI knowledge bases that can be quoted based only on the number of documents.
Also called "enterprise knowledge base", some quotes only include importing ready-made data into a tool; some also include data organization, job permissions, existing system access, real problem testing, formal deployment and subsequent maintenance. The name is the same, but the things they do may be completely different.
Therefore, after getting the quotation, don’t ask which one is cheaper. First check whether the following six tasks have been included:
- Inventory and organize data;
- retrieval, answering and provenance display;
- Position authority and sensitive content control;
- Access to existing data systems or business systems;
- Real problem testing, modification and acceptance;
- Models, cloud resources, software accounts and subsequent maintenance.
Only by writing these six items clearly can the total price be meaningful for comparison.
The first cost: whether the data can be used directly
A lot of information doesn’t necessarily mean a lot of work; a little information doesn’t necessarily mean it’s easy. What really affects the cost is whether the data can be directly used as the basis for the answer.
If the system, product descriptions and frequently asked questions have been divided into catalogs, the versions are clear, and someone is responsible for updating, the organizing work will be much less. If there are multiple versions of the same thing, the scans are not searchable, the table format is confusing, or no one knows which copy is still valid, you need to clean it up first.
The quotation should clearly state what data will be processed, what formats will be supported, whether scan identification will be included, how version conflicts will be handled, and who will be responsible for updating the data. Just write "Import all documents", it is easy to be confused during acceptance.
Second fee: Answers must be verified by real questions
After putting the information in, the system can answer the questions, but it can only be considered the beginning. For formal use, you also need to check whether the answers have found the correct information, whether the source can be shown, and whether there will be random answers when there is insufficient information.
This part of the cost usually comes from question set compilation, retrieval and answer rule adjustments, answer source display, error case handling, and multiple rounds of testing. The more complete the real issues provided by the company, the easier it is for both parties to see in advance how far they have to go.
The quotation cannot just say "connect to a certain model". The model is only part of it. How to find the information, how to limit the answers, and how to leave the test results determine whether this knowledge base can be accepted.
The third and fourth fees: permissions and system access
If all employees view the same public information, permissions will be simpler. If the content seen by sales, finance, human resources and management is different, the original accounts, departments and data permissions must be matched.
Data may also be stored in network disks, document platforms, OA, CRM or other internal systems. Whether it can be read through the interface, how often it is synchronized, how to invalidate it after deletion, and what to do if there is an interface error will all increase the work.
These two items are most likely to be omitted in low-price quotes. In the end, either everyone shares the same account, or they still have to manually upload data repeatedly. When making an inquiry, you should directly ask: what level of authority is required, which systems need to be connected, and who is responsible for the interface work.
The fifth and sixth expenses: acceptance and continued operation
Knowledge base acceptance cannot just look at a few questions asked during the demonstration. Authentic questions confirmed by both parties should be used to separately check the results for correct answers, incorrect information, insufficient permissions, missing information, and updated content.
After the project goes online, costs such as model calling, server, storage, scanning and identification, monitoring, and software accounts will also be incurred. Data continues to increase, business rules change, and someone needs to update and retest.
A clear quotation will list the project construction fees, third-party ongoing fees and subsequent maintenance separately, and explain who holds the account and who pays the fees. In this way, you won’t find out before the launch that there are a series of services that must be renewed in addition to the development fee.
How to finally compare three quotes
When comparing quotes in the knowledge base, you can put the content of each supplier into the same table, without comparing the total price first, and just answer the following questions item by item:
- which data is processed and which is not;
- which pages, functionality and management capabilities are delivered;
- What level of authority should be achieved and which existing systems should be connected?
- What real questions should be used for acceptance and how to record the results;
- Who pays for the model, server and software costs;
- Who will update the data after going online and how long will the maintenance last?
If an item is not written, mark it as "unconfirmed" first and don't assume it is included. After these spaces are filled in and you look at the total price, you can usually tell where the difference comes from.
Enterprise knowledge bases are not quoted by a chat box, but by a set of working methods that can consistently find reliable answers.
