Knowledge Base (Digital Employee)

A Digital Employee can associate one or more knowledge bases as its knowledge sources. Unlike workflow agents, Digital Employees do not passively execute a single lookup following preset retrieval strategies. Instead, they can autonomously decide what keywords to search with and whether to search multiple times, giving them stronger reasoning capabilities when facing complex questions.

Knowledge Base Configuration

The knowledge base configuration entry is located under the Knowledge Base menu item in the Digital Employee detail page. Configuration follows the same model as agent knowledge base strategy settings:

• Select Knowledge Bases: Check one or more created knowledge bases
• Retrieval Strategy: Configure library selection strategy (Select All / TopN / Minimum Score), retrieval document count, similarity threshold, and other parameters
• Post-Processing Strategy: Configure context enhancement, rerank model, and other post-processing parameters

For detailed configuration instructions, see Library Strategy Settings.

Autonomous Retrieval: The Core Difference

A workflow agent's knowledge base retrieval is deterministic: when the workflow reaches a retrieval node, it executes one search using preset keywords and strategy, returns results, and continues to subsequent nodes.

A Digital Employee is fundamentally different. It has no fixed retrieval node—retrieval behavior is part of the AI's autonomous decision-making.

Autonomous Keyword Selection

When a user asks the Digital Employee a question, the Digital Employee analyzes the question's semantics and constructs its own search keywords. These keywords are not necessarily the same as the user's original question. For example:

• User asks: "Any updates on that financial analysis proposal we discussed last time?"
• The Digital Employee might construct keywords: "financial analysis proposal progress", and may even try different keyword combinations multiple times

This autonomy avoids the common "keyword rigidity" problem in workflows—if preset keywords are inaccurate, the workflow can only return irrelevant results. Digital Employees can adjust keywords and re-search when initial results are unsatisfactory.

Autonomous Retrieval Count

More critically, Digital Employees can perform multi-round retrieval. When initial retrieval results cannot fully answer the user's question, the Digital Employee doesn't give up. Instead, it constructs a second round, then a third round of retrieval based on the information already obtained.

This multi-round retrieval capability is particularly valuable in these scenarios:

• Progressive Questions: User asks "How is technology A being applied?" → Retrieve that A is applied in a project → Then retrieve more details about that project
• Cross-Validation: Retrieve information on the same topic from different knowledge bases for cross-comparison
• Information Completion: First round found architecture description but lacks performance data → Second round targets performance metrics

Synergy with Knowledge Graph

The Digital Employee's autonomous multi-round retrieval capability is most effective when paired with knowledge graphs. Knowledge graphs structure document entities and relationships into graph networks, which are naturally suited for multi-hop reasoning.

Specifically, when a Digital Employee retrieves an entity node from a knowledge graph, it not only gets the node's information but also sees connected neighboring nodes. Based on these relationships, the Digital Employee can autonomously decide "what to look up next":

1. User asks: "What is the safety profile of this drug for children?"
2. Digital Employee searches the graph, finds "Drug A" node, connected to "Clinical Trial" node
3. Digital Employee autonomously initiates a second retrieval, querying the "children's group" data within that clinical trial
4. Based on children's group results, retrieves "adverse reactions" related nodes

This process is not a pre-orchestrated workflow but a dynamic multi-hop retrieval path autonomously planned by the Digital Employee based on the knowledge graph's structure. Each retrieval decision builds on the previous round's results, forming a dynamic reasoning chain.

This is fundamentally different from the traditional RAG model of "single retrieval + LLM answer": traditional RAG treats retrieval as "looking up materials"; Digital Employee + Knowledge Graph treats retrieval as "conducting research".

Configuration Recommendations

• Multi-hop reasoning scenarios (e.g. regulatory interpretation, academic research) strongly recommend associating knowledge graph-type knowledge bases
• When configuring multiple knowledge bases, the Digital Employee can autonomously decide which library to query and can perform cross-library associative retrieval
• Retrieval parameters should not be set too strictly (e.g. excessively high similarity thresholds), as Digital Employees have autonomous judgment and can self-filter and dig deeper after initial retrieval