AI Chat
In the Chat tab of the Knowledge detail screen, you can have conversations with AI based on collected document data. This is not a simple chatbot — it works by searching the Knowledge you've accumulated and generating answers based on the results. Therefore, you receive accurate responses grounded in actually collected documents, not answers fabricated by the AI.
This feature is implemented through a RAG (Retrieval-Augmented Generation) pipeline. When a user enters a question, relevant document chunks are first retrieved (Retrieval), and then the retrieved context is passed to the LLM to generate an answer (Generation). Each answer displays the referenced source information so you can directly verify and validate the basis of the response.
Screen Structure
The AI Chat screen consists of three panels.
Left Panel — Conversation Session List
The area for managing previous conversation history.
| Feature | Description |
|---|---|
| New Conversation | Click the + New Conversation button at the top to start a new session. |
| Search | Search previous sessions by conversation title in the search field at the top of the session list. |
| Date Grouping | Sessions are automatically grouped by Today, Yesterday, Last 7 Days, Last 30 Days, etc. |
| Delete Session | Individual sessions can be deleted from the context menu of each session item. |
Center Area — Chat Messages
The core area where the actual conversation takes place.
- User Messages: Your questions are displayed as message bubbles.
- AI Responses: Answers generated based on retrieved document context are rendered in markdown format.
- Citation Pills: Referenced sources within answers are displayed as numbered pills. Click them to view the original text.
- Input Field: Write and send questions in the text input area at the bottom of the screen. Press
Enterto send, andShift + Enterfor line breaks.
Right Panel — Settings and Source Filter
The area for fine-tuning chat behavior including RAG strategy, LLM settings, and source filtering. Each setting is described in detail in the sections below.
RAG Strategy
The RAG strategy determines how related documents are retrieved for the user's question. Selecting the optimal strategy based on the nature of the question can significantly improve answer quality.
| Strategy | Description | Recommended Use Case |
|---|---|---|
| vector | Semantic similarity search based on embedding vectors. Finds chunks most semantically similar to the question from the Vector DB. | General queries, conceptual questions |
| text | Keyword full-text search based on the BM25 algorithm. Strong at precise keyword matching. | Exact terms, proper nouns, code names |
| graph | Graph DB exploration based on GraphCypherQAChain. Reasons about relationships between entities to generate answers. | Entity relationship queries (e.g., "What is the relationship between A and B?") |
| hybrid | Performs vector search and text search in parallel, then merges results using the Reciprocal Rank Fusion (RRF) algorithm. Leverages the advantages of both search methods. | Recommended for most cases (default) |
| agentic | LangGraph-based agentic RAG. The LLM autonomously performs query rewriting, document relevance evaluation, and multi-step retrieval. If retrieval results are insufficient, it rewrites the question and searches again (up to 3 times). | Complex multi-step questions, ambiguous queries |
In most cases, the hybrid strategy provides the most balanced results. Use text when you need to find specific terms or proper nouns precisely, and vector for open-ended questions where semantic similarity matters. The agentic strategy is effective for complex questions requiring multi-step reasoning, but response times may be longer.
To use the graph strategy, GRAPH must be included in the storage targets when creating the Knowledge. The graph strategy cannot be selected for Knowledges created without the GRAPH target.
LLM Settings
In the right panel, you can configure the LLM model and parameters used for generating AI responses.
Model Selection
Available LLM models are displayed in a dropdown, grouped by Provider.
| Provider | Representative Models |
|---|---|
| OpenAI | GPT-4o, GPT-4o-mini, etc. |
| Claude | Claude 3.5 Sonnet, Claude 3 Haiku, etc. |
The list of available models may vary depending on the system administrator's configuration. If a Provider is not displayed, contact the administrator to verify API key settings.
Temperature Control
Temperature is a parameter that controls the randomness (creativity) of AI responses.
| Range | Characteristics | Best For |
|---|---|---|
| 0.0 ~ 0.3 | Deterministic, consistent responses | Fact-based queries, accurate information verification |
| 0.4 ~ 0.6 | Balanced responses | General Q&A |
| 0.7 ~ 1.0 | Diverse and creative responses | Brainstorming, summary generation |
| 1.0 ~ 2.0 | Very high randomness | Experimental use |
The default value is 0.7.
For accurate fact-based answers, lower the temperature to 0.1~0.3. For creative answers or diverse perspectives, raise it to 0.7 or higher. For RAG-based chat where document-grounded answers are important, the 0.3~0.5 range is generally recommended.
Source Filtering
In the Source Filter section of the right panel, you can limit the scope of documents that the AI references.
How It Works
- Select All (Default): If no filter is set, all documents collected in the Knowledge are searched.
- Select Specific Documents: Check the desired documents, and only chunks from those documents will be included in the search scope.
Usage Scenarios
| Scenario | Setting |
|---|---|
| Search comprehensively across entire Knowledge | No filter (select all) |
| Ask questions about a specific report only | Check only that report document |
| Get answers based only on recent documents | Check only recently collected documents |
| Limit to a specific source type | Check only desired types such as web crawling/file upload |
The list of selected document IDs is passed to the search engine as the document_ids parameter. This filter is applied equally to both vector search and text search, so chunks outside the selected document scope are not included in search results.
Citation
To ensure the reliability of AI responses, sources used in answers are displayed as citation pills.
Citation Display Format
Numbered pills such as [1], [2] are displayed inline within the AI response text. Each number points to a document chunk referenced in generating the answer.
Viewing Sources
Clicking a citation pill opens a Source Detail Modal where you can view the following information:
| Item | Description |
|---|---|
| Document Name | The name of the original source document |
| Section | The section the chunk belongs to within the document (if available) |
| Chunk Number | The sequence number of the chunk within the document |
| Original Content | The full text of the actually referenced chunk |
| Similarity Score | Relevance score to the question (%) |
Using Citations
- You can directly verify answer accuracy by cross-referencing with the original text.
- When multiple sources are cited, you can compare each source to check information consistency.
- You can craft more specific follow-up questions based on the source documents.
Making a habit of checking cited sources is the most effective way to prevent AI hallucination. Always verify the original text when using it for important decision-making.
Conversation Management
Session Storage Method
Conversation sessions are stored in browser local storage. Since conversation history is not stored on the server (stateless), clearing browser data will also delete conversation records.
Automatic Session Titling
When starting a new conversation, the first 50 characters of the first message are automatically set as the session title. This makes it easy to identify conversations in the session list.
Session Limit
| Item | Value |
|---|---|
| Max Sessions | 50 |
| Behavior on Overflow | The oldest session is automatically deleted |
Since conversation history is stored only in browser local storage, all conversation records will be lost when changing browsers, using incognito mode, or clearing cache. For important conversation content, use the copy/share features in response actions to save them separately.
Response Actions
Various action buttons are provided at the bottom of each AI response message.
| Action | Description |
|---|---|
| Copy | Copies the AI response text to the clipboard. |
| Share | Exports the response in markdown format. Formatting is preserved when pasting into other tools. |
| Like / Dislike | Provides feedback on response quality. Used for future quality improvement. |
| Save as Note | Saves the response content as a note for later reference. |
| Regenerate | Generates a new answer for the same question. Useful when you want a different perspective. |
| Stop | Immediately stops streaming during response generation. Used to halt unnecessary responses early. |
The same RAG strategy and LLM settings are applied during regeneration, but different answers may be generated depending on the LLM's temperature value. The closer the temperature is to 0, the more similar the regenerated result will be to the original answer.
Tips for Writing Effective Questions
Tips for getting better answers from AI Chat.
| Method | Example |
|---|---|
| Ask specifically | ❌ "Summarize this document" → ✅ "Tell me the 2024 revenue target from this document" |
| Provide context | ❌ "What's the procedure?" → ✅ "Explain the new employee onboarding procedure step by step" |
| Specify output format | "Organize the key points in bullet list format" |
| Use source filters | Selecting only relevant documents narrows the search scope and improves accuracy |
Next Steps
- Search Test — Verify Knowledge quality by checking pure search results without LLM
- Chunking and Options — Understand how chunking strategies and embedding models affect search quality
- Settings — Manage Knowledge storage targets and embedding models