Knowledge is a collection of business knowledge and rules provided for skills to use. By binding a knowledge library to an agent, you can standardize the agent's business calibers during data queries or report generation, ensuring that analysis results comply with real business rules.
The Knowledge module is used to accumulate the content, such as business rules, terms, and logic, within an enterprise, enabling an agent to follow unified business standards when skills are invoked. This can improve the accuracy of intent recognition and analysis results.
Taking how to add a knowledge library for chain operation analysis as an example, the following figure shows the effect.
(1) Choose Admin Console > Knowledge, and click Create Knowledge Library in the top right corner. In the Create Knowledge Library pop-up box, enter the following basic information and click OK to complete the knowledge library creation.
Name: Customize the knowledge library name to clearly reflect the business scenario.
Description: Briefly describe the scope and purpose of the knowledge in the knowledge library.
Icon: Set a preset icon to quickly identify the knowledge library.
(2) After the knowledge library is created, enter the knowledge library configuration page, where you can configure business rules and file knowledge.
Business rules are used to enter content such as enterprise business terms, classification standards, statistical calibers, abbreviation definitions, and common-sense definitions to help the agent understand the business context and standardize its analysis and response behavior.
Example
Core Product = Product with At Least $10 Million in the Annual Sales Amount (Including Model A and Model B)
Four Major Regions = Southeastern United States, Northeastern United States, Midwestern United States, and Southwestern United States.
High-Value Customer = Customer with At Least $5000 Per Order and At Least 3 Annual Consumptions
To configure business rules more efficiently and improve system recognition and execution efficiency, you can refer to the following recommended configuration standards for clearer rule expressions and semantics, reducing system parsing deviations.
A fixed sentence pattern is used. (If the xx condition is met, then execute xx.)
If a [specific condition] is met, select/add [target content] directly.
When a [specific business term] is queried, add/select/query [target content].
Condition descriptions must be associated with specific data items.
Condition descriptions must specify the exact data items. Vague expressions such as related data and business-related must be prohibited. Instead of intra-enterprise jargon, standard names recognizable by the system need to be used uniformly.
If multiple data items need to be associated, the logical relationship (for example, "appearing simultaneously" or "containing at least one") needs to be clearly stated.
Executions must be positively phrased.
For the execution description in the latter part of a rule, the "specific execution to be performed" needs to be described in a positive tone, such as "Select the XX subject.", "Add the XX filter condition.", and "Query the XX data.", rather than in a negative tone (for example, "Do not select the XX subject.").
These standards are recommended configuration standards and may not apply to all LLMs. Due to the differences in semantic parsing capabilities and adaptation logic across LLMs used by different users, some rules may not take effect after configuration.
You can upload text files as knowledge sources. Currently, the following four formats are compatible:
DOCX
TXT
PDF
Markdown
On the knowledge library configuration page, you can perform daily management of the entered content:
Business rules: They can be searched, edited, and deleted.
File knowledge: Uploaded files can be searched, downloaded, and deleted.
The page allows you to intuitively view the number of used/remaining tokens of the current knowledge library and monitor the usage in real time.
The following operations are supported for an existing knowledge library:
Search: Enter the knowledge library name/description in the top search box to quickly locate the target knowledge library.
Rename: Click ... in the bottom right corner of a knowledge library, select Rename, and rename the knowledge library.
Delete: Click ... in the bottom right corner of a knowledge library, select Delete, and delete the knowledge library that is no longer in use. (The deletion is irreversible.)
(1) The agent can recognize only plain text content in the file knowledge. Non-text information, such as images and charts, in the file cannot be parsed and used by the agent.
(2) The upper token capacity limit for a single knowledge library is 4,000 tokens. If the upper limit is exceeded, you will not be able to add more rules or upload more files.
(3) The rule effectiveness is subject to the LLM capabilities. If some rules do not take effect, you can fine-tune the rules and test them again.