LongbridgeAI Docs

English

简体中文
繁體中文

English

简体中文
繁體中文

Theme

Knowledge Base

What It Is ​

Upload your local documents to the platform, where they are parsed, chunked, and vectorized for the Agent to retrieve — so the Agent answers based on your private knowledge instead of relying only on the model's general knowledge. Typical scenarios: product FAQ customer service, internal policy Q&A, and research material analysis.

Knowledge base home page: the "Create Knowledge Base" entry in empty state

Creation Flow (Four Steps) ​

  1. Create a knowledge base: fill in the name and description, and choose the retrieval settings. Three retrieval methods are supported:

    • Hybrid retrieval (recommended default): vector + full-text combined
    • Vector retrieval: semantic similarity matching, suited for conversational questions
    • Full-text retrieval: exact keyword matching, suited for terminology, code, and ID-style queries
    • You can also click "Create Blank Knowledge Base" to create the base first and add files later

    Create knowledge base: name, description, and retrieval settings (hybrid retrieval is the recommended default)

  2. Upload files: drag and drop or select files, up to 5 per batch and 10MB each; supports DOCX, PPTX, HTML, PDF, MD, CSV, XLSX, VTT, JPG, PNG, TXT

    Upload files: drag-and-drop or select, with supported formats and size limits

    Uploaded file list: a large document can be split into multiple files and uploaded in batches (example: a book split into 4 PDFs by page range)

  3. Text settings: customize the chunk length per file type (default 1024 characters), text preprocessing rules (replace consecutive spaces/line breaks, remove URLs and emails), semantic chunk enhancement (enable image content understanding via vlm), and click "Preview Chunks" to preview the chunking result — chunking quality directly determines retrieval quality, so this step is worth extra time

    Text settings: chunk length, preprocessing rules, semantic enhancement, and chunk preview

  4. Process and finish: view the embedding progress (each file shows its own steps and percentage), with a summary below of this batch's segmentation mode, chunk length, preprocessing rules, and retrieval settings; you can go straight to the document list page while processing continues in the background

    Process and finish: embedding progress and configuration summary

Day-to-Day Management ​

  • Knowledge base list: the knowledge base tab shows all knowledge bases as cards (retrieval method, document count, update time), with a storage usage progress bar at the top; the "…" in the top-right corner of each card opens management actions

    Knowledge base list page: storage progress bar and knowledge base cards

  • Document list (knowledge base detail page): filter by processing status and search by name; shows character count, retrieval count (which documents are actually used), upload time, and status (such as "Indexing"); each document has an enable/disable toggle and more actions, and you can keep adding files via "Add File" in the top right

    Knowledge base detail page: document list with status and enable toggles

  • Knowledge base settings: modify the knowledge base's name, description, and retrieval settings (switch between hybrid/vector/full-text at any time; hybrid retrieval offers advanced settings to adjust weights)

    Knowledge base settings page: modify name, description, and retrieval settings

  • Chunk editing: open a single document to view the final chunks handed to the LLM; you can re-edit specific chunks and their keywords; "Add Chunk" inserts custom content with keywords at the top of the document

  • Recall testing: see the next section

Recall Testing ​

The entry is in the top-right corner of the knowledge base detail page (to the left of the Add File button). Enter the source text on the left (up to 200 characters), confirm the retrieval method, and click "Test":

Recall test: source text input, retrieval method, and test button

The right side shows the matched recalled chunks: each chunk displays its index, character count, and relevance score (e.g. 0.67), used to judge retrieval quality:

Recall test result: 3 recalled chunks with chunk info and scores

  • Each recalled chunk is labeled with its source file at the bottom, so you can confirm which document the content came from
  • The "Records" list at the bottom left keeps past tests (retrieval method, query, time), making it easy to compare results for the same query across retrieval methods
  • Use "questions real users would actually ask" as queries rather than sentences copied verbatim from the documents — verbatim sentences naturally score high and won't reveal real-world performance

Recall test records: query history list, recalled chunks labeled with source file

Using It in a Flow ​

Two mounting points; in both, the Agent automatically retrieves from the selected knowledge bases when needed:

① Agentic Chat: in the "Knowledge Base" section of the Agent configuration page, click "+ Add Knowledge Base" to mount one (adjacent to the Skills section; see the configuration page screenshot in Skill System).

② Agent node in Chatflow / Workflow: turn on the "Knowledge Base" switch in the node editing panel and click + to select the knowledge bases to mount:

Agent node: the selection entry appears after the knowledge base switch is turned on

Selected knowledge bases are shown as tags (multiple can be mounted):

Agent node: "Investment Books Knowledge Base" mounted

Write Good Descriptions to Help the Agent Retrieve ​

The Agent relies on the knowledge base's name and description to decide "which questions require this base" — the description should clearly state the scope of content covered (e.g. "Investment knowledge base, currently including options-related data"), so the Agent knows to retrieve from it when an options question comes in; a vague or missing description leads to missed retrievals when they should happen:

Knowledge base settings: a description that clearly states the content scope helps the Agent decide when to retrieve

Retrieval in Action at Runtime ​

Once configured, when a user's question falls within the knowledge base's coverage, the Agent retrieves automatically: the reasoning process shows a "Knowledge Base Retrieval" step (including the search query), and citations in the answer body are marked with Document +N tags:

Asking an options question in chat: the reasoning process shows the knowledge base retrieval step, and the answer carries Document citation tags

Click the "Knowledge Base Retrieval" step to view the request/response details — the request contains the target knowledge bases and search query, and the response shows each matched chunk's content, relevance score, and rank, useful for troubleshooting retrieval quality:

Knowledge base retrieval details: request (dataset_uids/query) and response (chunk content/score/rank)

The prompt should explicitly state "answer only based on retrieved content, and say so honestly when nothing is retrieved" to prevent the model from fabricating answers.

Notes ​

  1. Storage scales with plan: Starter 50MB / Pro 5GB / Premium 20GB; once over the limit, new uploads are blocked, and over-limit files are marked disabled starting from the oldest by upload time (excluded from retrieval); delete files or upgrade to restore — see Workspace & Plans
  2. Embedding billing: document vectorization consumes usage quota by token
  3. Compliance: content in customer-facing knowledge bases must come from compliant sources and must not contain PII or undisclosed material information; see Data Usage & Privacy Restrictions
  4. Before going live, run recall tests on typical questions to confirm retrieval quality before publishing