Chat Assistant Configuration#

A chat assistant (called a Dialog in the backend) is the central configuration unit that binds an LLM, a set of knowledge bases, retrieval parameters, and behavioral feature flags together into a deployable chat interface.

Data Model#

The Dialog class in api/db/db_models.py holds all persistent configuration. Key fields:

The default prompt_config value is 1:

{
  "system": "",
  "prologue": "Hi! I'm your assistant. What can I do for you?",
  "parameters": [],
  "empty_response": "Sorry! No relevant content was found in the knowledge base!"
}

prompt_config Feature Flags#

All behavioral toggles live inside prompt_config (a JSON blob on the Dialog record). The frontend schema in use-chat-setting-schema.tsx documents every valid key:

At runtime in dialog_service.py, these flags are read directly from dialog.prompt_config using .get() with safe defaults (e.g., prompt_config.get("quote", True), prompt_config.get("keyword", False)).

Frontend Validation (Zod Schema)#

The form schema in use-chat-setting-schema.tsx composes sub-schemas for each concern:

• prompt_config — validated by promptConfigSchema; system is the only required string field (min length 1) 2
• name — required, min length 1 3
• dataset_ids — array (may be empty) 4
• llm_setting — validated by LlmSettingFieldSchema from the shared LLM settings component 5
• Retrieval sliders — vectorSimilarityWeightSchema, similarityThresholdSchema, topnSchema from shared components 6
• Rerank — rerankFormSchema 7
• Metadata filter — MetadataFilterSchema 8

Web Search via Tavily#

Tavily integration adds real-time web results to the retrieval context. Two things must both be true for web search to fire:

1. prompt_config["tavily_api_key"] must be non-empty
2. The internet flag in the request must be truthy

The backend helper _should_use_web_search() enforces this — if either condition fails, it returns False. The internet flag accepts bool, int, float, or string ("true", "1", "yes", "on", etc.) via _normalize_internet_flag().

When enabled, Tavily.retrieve_chunks() is called with search_depth="advanced" and max_results=6 9. Results are converted to the standard chunk format (with vector_similarity fixed at 1.0 and term_similarity at 0) and merged directly into kbinfos["chunks"] alongside knowledge base results 10.

Retrieval Flow (Simplified)#

async_chat()
    │
    ├─ SQL path? (if field_map from KB) → use_sql()
    │
    └─ Vector path:
         ├─ retriever.retrieval() ← top_n, similarity_threshold, vector_similarity_weight, top_k
         ├─ toc_enhance? → retriever.retrieval_by_toc()
         ├─ Tavily (internet + tavily_api_key)? → extend chunks
         └─ use_kg? → kg_retriever.retrieval()

See dialog_service.py lines 757–790 for the complete branching logic.

Key Source Files#