社内ナレッジ検索アプリ

Overview

A RAG-type web application that answers based solely on information existing in internal documents, with a verifiable flow from the answer back to the relevant source material. Rather than a chat interface, it is designed around a verification workflow of question → search → evidence extraction → answer → citation → source confirmation.

Main Features

• Integrated AI Answers and Citations - Answers include document name, page, heading, and excerpt, with citations enabling navigation to the relevant location in Document Viewer
• Access Control - Evaluates viewing department before search and excludes documents outside access scope
• Hybrid Search - Combines vector search and keyword search
• Document Version Management - Prioritizes newer versions within the same document family and excludes older versions from normal search
• Flexible Filtering - Specify search conditions by category, department, document type, and date
• Evidence Visualization - Confirm search query, retrieved chunks, similarity scores, and answer generation stages in Developer View
• Multi-format Support - Document registration for PDF, DOCX, TXT, Markdown, CSV (maximum 8MB)
• Follow-up Support - Supplements previous conversation into search query and re-searches each time

Characteristics

• Returns "Cannot be confirmed from registered materials" when there is insufficient evidence
• As a hallucination countermeasure, prohibits answers and speculation outside of context
• As a prompt injection countermeasure, excludes imperative text from answer sources
• API keys are managed as server environment variables and not included on the client side
• Demo mode does not send data to external AI