Knowledge Base

The Knowledge Base is LanSphere’s hosting and retrieval capability for enterprise materials. Upload policy documents, product manuals, technical docs, and more into a Knowledge Base, then associate it in a Workflow, Chatflow, or Agent — so your app answers from your organization’s real materials rather than the LLM’s generic memory.

This article covers how a Knowledge Base works, the two build modes, the full standard-creation steps, indexing and retrieval settings, the Retrieval Test, document management, Third-party Knowledge Base integration, and permission management. It is for developers and admins building the enterprise knowledge foundation.

How a Knowledge Base Works

A Knowledge Base uses retrieval-augmented generation (RAG). Answering a question goes through three steps:

1. Retrieve: Find the content chunks most relevant to the user’s question by semantics or keywords.
2. Augment: Provide the retrieved chunks as context to the LLM together with the question.
3. Generate: The LLM composes the answer from these real materials instead of answering from thin air.

So the quality of the materials, the soundness of chunking, and the retrieval configuration directly determine how accurate and trustworthy the app’s answers are.

Two Build Modes

When creating a Knowledge Base, you choose between two build modes:

Tip: First time here? Start with Simple mode. The Knowledge Pipeline targets users who need a custom processing flow — just follow the on-page prompts.

Standard Creation in Five Steps

Go to “Knowledge Base” in the left menu, choose standard creation, and complete the following five steps.

Step 1: Basic Info

• Name: Required, up to 30 characters.
• Description: Optional, up to 300 characters. State the material scope and purpose so teammates can understand and pick it.

Step 2: Upload Files

Supported formats include DOCX, TXT, MD, XLSX, XLS, CSV, PDF, and more, up to 15MB per file.

Step 3: Chunking Strategy

Chunking determines how documents are split into retrievable chunks. The platform provides 5 preset strategies:

There are three chunking modes:

• Rule-based Chunking: Splits documents by delimiter and length rules.
• Q&A Chunking: Organizes materials into question-answer pairs — for high-precision Q&A scenarios.
• Parent-Child Chunking: Determines the parent chunk at the “Paragraph” or “Full Text” level first, then splits child chunks; retrieval hits child chunks while generation brings back the parent context — precise yet complete.

Preprocessing options: all preset strategies enable “merge consecutive whitespace” by default; “remove URLs and email addresses” is optional for cleaning out noise.

Step 4: Parameter Confirmation

• Indexing method: “Vector Indexing” (default) or “Keyword Indexing.” With Parent-Child Chunking or Q&A Chunking, indexing is locked to Vector Indexing.
• Embedding Model: Required with Vector Indexing — vectorizes the text chunks.
• Rerank Model: Optional. Once configured, retrieval results are re-ranked for better precision.

Note: Once Vector Indexing is chosen, you cannot switch back to Keyword Indexing. Confirm the indexing method before creating.

Step 5: Confirm Creation

After confirmation, the system processes files one by one and the page shows each file’s progress; if an individual file fails, you can reprocess that file. Once everything succeeds, click “Enter Knowledge Base” to start using it.

Adding More Files

When appending files to an existing Knowledge Base, new files automatically inherit the chunking mode from creation — locked and unchangeable. To change the chunking approach, create a new Knowledge Base.

Retrieval Settings

A Knowledge Base supports three retrieval methods:

• Vector retrieval: Recall by semantic similarity — for natural-language questions.
• Full-text retrieval: Recall by keyword match.
• Hybrid retrieval: Weighted combination of vector and full-text — both semantics and exact hits.

Other key parameters:

• Top K: Number of chunks returned per retrieval; default 3.
• Score threshold: Filters out low-score chunks by match score; off by default.
• Rerank Model: Optional; re-ranks retrieved results for better final hit precision.

Retrieval Test

After building a Knowledge Base, verify it with the “Retrieval Test”: enter test questions and inspect the retrieved chunks, match scores, and source documents. If retrieval is unsatisfying, adjust the retrieval method, Top K, or chunking strategy and test again.

Document Management

On the Knowledge Base details page, manage the onboarded documents:

• Document list: View all documents and their status.
• Enable / disable: Temporarily take a document out of retrieval without deleting it.
• Chunk viewing and management: Inspect how a document was split.
• Metadata tagging: Add tagging info to documents.
• Download / rename / delete: Maintain the documents themselves.

Third-party Knowledge Base

If your organization’s knowledge materials are already hosted in another system, connect it over API as a “Third-party Knowledge Base” and use it directly — no data migration into LanSphere needed.

Permission Management

Knowledge Base permissions follow the same rules as app permissions:

• “Only me” by default; can be opened to “selected team members” — multi-select and select-all supported.
• Authorized members can use the Knowledge Base but cannot delete documents; settings are view-only.
• Ownership transfer supported: operate in the “Permissions” pop-up; after transfer, the original Owner loses all permissions while collaborators and visible scope stay unchanged.

For full rules, see “Members & Permissions.”

List Filtering

The Knowledge Base list can be filtered by “build mode” (Simple / Expert), and “Created by Me” quickly locates the Knowledge Bases you created.