NeuCharFramework (NCF)
  • NCF - NeuCharFramework
  • Projects

    • Preparation
    • NcfPackageSources Source Guide
    • DynamicWebApiEngine
    • Deployment
    • MCP (Model Context Protocol) Module
    • Senparc.AI
  • AIKernel Local Fine-Tuning
  • Help

    • Online Resources
    • Q&A Community
    • QQ Group (147054579)
    • Senparc WeChat SDK
  • Gitee
  • GitHub
  • English
  • 简体中文
GitHub
  • NCF - NeuCharFramework
  • Projects

    • Preparation
    • NcfPackageSources Source Guide
    • DynamicWebApiEngine
    • Deployment
    • MCP (Model Context Protocol) Module
    • Senparc.AI
  • AIKernel Local Fine-Tuning
  • Help

    • Online Resources
    • Q&A Community
    • QQ Group (147054579)
    • Senparc WeChat SDK
  • Gitee
  • GitHub
  • English
  • 简体中文
GitHub
  • Source Development

    • NcfPackageSources Beginner Quickstart (60 Minutes)
    • NcfPackageSources Source Guide
    • Project Relationships, Synchronization, and Release
    • Version Upgrade Notes
    • NcfPackageSources_Include_NcfSimulatedSite.sln Project Map
  • XNCF Source Analysis

    • NCF Capability Source Deep Dive (Practical)
    • XNCF Extension Library Guide (Senparc.Xncf.xxx)
    • XNCF Module Documentation Map
    • Senparc.Web and Admin
    • XncfBuilder Isolated Development: Source Analysis
  • Admin Feature Guides

    • Admin help directory and short notes
    • Admin logs and diagnostics
    • AdminChat conversations, Harness and attachments
    • AdminChat Memory and Conversation Settings
    • NeuBell outbound WebHook management
    • NeuCharPivot boards, aggregation and access
    • MCP Manager: published services and client connections
    • AIKernel models, vectors and usage
    • AIKernel Local Fine-Tuning
    • KnowledgeBase ingestion and recall tests
    • PromptRange ranges, testing and model settings
    • Skill import, testing and publication
    • FileManager assets and upload policies
    • AgentsManager agent management
    • AgentsManager 3D Status View
    • AgentsManager: HIL and NeuChar Workflow Integration
    • NeuChar Workflow: Operations and {{= ... }} Expressions
    • Workspace and AdminChat
    • Sandbox session management and environment setup
    • DynamicData tables, layouts and pages
    • SenMapic crawl tasks and results
    • Weixin official accounts and followers
    • Personal Weixin Bot and Admin integration
  • NCF Library Source Analysis

    • Senparc.Ncf.Core
    • Senparc.Ncf.Database (Database Foundation)
    • Senparc.Ncf.Repository
    • Senparc.Ncf.Service
    • Senparc.Ncf.SMS
    • Senparc.Ncf.Mvc.UI
    • Senparc.Ncf.Log
    • Senparc.Ncf.Utility
    • Senparc.Ncf.XncfBase
    • Senparc.Ncf.AreaBase
    • Senparc.Ncf.DatabasePlant
  • Core Interface Source Analysis

    • IXncfRegister Interface (Current)

KnowledgeBase ingestion and recall tests

Knowledge bases and vectorization

  1. Create a knowledge base and choose its embedding model and vector storage.
  2. Associate the intended source material in its configuration.
  3. Run Vectorize, inspect the result, and resolve failures before testing.
  4. Run recall tests using realistic questions and inspect source text.

Editing source files or changing an embedding configuration does not by itself prove the stored vectors are current. Rebuild as required. Keep source access permissions and retention requirements in mind before sending content to an external embedding provider.

Create a knowledge base

New knowledge base opens the name and model configuration form. Choose an embedding model, vector database and any required chat model. Embeddings support retrieval; the chat model generates answers and is not a substitute. After saving the definition, associate sources through Configuration and run Vectorize. Creating or saving the definition does not finish the entire ingestion process.

Recall tests and Top K

Select a vectorized knowledge base, enter a realistic question, and inspect returned snippets, source names, links, and ordering. View full text opens snippet details. History shows the question, knowledge base, snippet count, highest similarity and test time.

Similarity scores are ranking signals, not recall percentages or correctness probabilities. Different models/backends can use different scales. If reranking is enabled, distinguish its score and order from the original vector ranking. Inspect full paragraphs before drawing conclusions. A single successful query does not establish the quality of the entire knowledge base.

Returned snippet count (Top K)

The slider and number input under Retrieval settings limit snippets returned by subsequent tests on this page. The range is 1–20, with 5 suggested as a starting point. Increase for more coverage or decrease to reduce noise. Top K is not a similarity threshold, recall rate or document count. Changing it does not re-vectorize the knowledge base or alter global model parameters.

Edit this page on GitHub
Last Updated:
Contributors: JeffreySu
Prev
AIKernel Local Fine-Tuning
Next
PromptRange ranges, testing and model settings