Conversation Auto-Naming#

Dify automatically generates a concise title for every new chat conversation using an LLM call against the first user message. There are two distinct entry points into this system:

1. Automatic naming at message time — triggered inline during the message pipeline, in a background daemon thread.
2. On-demand naming via API — triggered explicitly by a POST /conversations/<id>/name request with auto_generate: true.

Both paths ultimately call the same generator, but they go through different code paths and have different validation rules.

How It Works#

1. Background auto-naming at message time#

MessageCycleManager.generate_conversation_name() fires a 1-second delayed Timer thread on the first message of any non-completion conversation. It checks extras["auto_generate_conversation_name"] (defaults to True) before spawning 1.

The worker thread, _generate_conversation_name_worker(), adds a Redis cache keyed on conv_name:{conversation_id}:{md5(query)[:16]} (TTL 1 hour) to avoid redundant LLM calls on retries 2. On LLM failure it falls back to query[:47] + "..." 3.

2. On-demand rename API#

Three API surfaces expose POST .../conversations/<id>/name:

All three validate with ConversationRenamePayload and delegate to ConversationService.rename().

Core Generator: LLMGenerator.generate_conversation_name()#

Full implementation. Key behaviors:

• Uses CONVERSATION_TITLE_PROMPT 4, which instructs the LLM to decompose input into Intention + Subject and return JSON with keys Language Type, Your Reasoning, and Your Output.
• Truncates queries > 2,000 characters to query[:300] + "...[TRUNCATED]..." + query[-300:] 5.
• Uses the tenant's default LLM via ModelManager.for_tenant().get_default_model_instance() with max_tokens=500, temperature=1 6.
• Parses JSON response; if "Your Output" is absent/blank, falls back to the raw query 7.
• Caps the resulting name at 75 characters (truncating with "...") 8.
• Emits a GENERATE_NAME_TRACE trace task for observability 9.

Validation Gap (FIXED)#

The manual naming path has an explicit non-blank guard. In ConversationRenamePayload, the validate_name_requirement model validator rejects name when it is None or whitespace-only (when auto_generate=False). The JSON schema also enforces a non-blank pattern via "pattern": r".*\S.*" 10.

The auto-naming path previously did not share this protection. In ConversationService.auto_generate_name(), the LLM call is wrapped in contextlib.suppress(Exception), and the result is assigned directly to conversation.name = name:

with contextlib.suppress(Exception):
    name = LLMGenerator.generate_conversation_name(...)
    conversation.name = name

11

Within LLMGenerator.generate_conversation_name() itself, the original behavior checked if answer == "": return "" — returning an empty string when the LLM produced an empty response before JSON parsing. This allowed blank or whitespace-only titles to be persisted to the database through the auto-naming path and the background worker path in _generate_conversation_name_worker() 13.

This has been fixed. The empty check now uses if not answer.strip(): and falls back to answer = query instead of returning an empty string. Empty or whitespace-only LLM responses now result in the original user query being used as the conversation title, preventing blank names from propagating through the system.

Key Files#