LLM Chat History
The Chat History input on PixelFlow's LLM nodes is not yet carrying conversation context between turns. What the connection does today, and what to use instead.
Not carrying context yet. You can wire one LLM node's output into another's Chat History input, the edge saves, and the receiving node reports Chat History · Connected — but the earlier turn is not currently included in the request sent to the model. A second node asked "what did you just suggest?" answers that it can't see any previous suggestion. Treat everything below as the intended design, not current behaviour, and see what to do instead.
Chat history is intended to let LLM nodes remember previous turns: connect one LLM node's output to the next node's Chat History input and the conversation context carries forward — the foundation for chatbots, agents, and any multi-turn dialogue.
Screenshot coming soon
A two-turn conversation as a graph
/images/pixelflow/guides/llm-chat-history.webp
How it is meant to work
Language model nodes expose a Chat History input alongside the prompt:
- First turn — the first LLM node takes your prompt and generates a response. No history is connected.
- Carrying context — connect that node's output to the next LLM node's Chat History input. The full exchange (prompt + response) travels with it.
- Subsequent turns — each node in the chain receives the entire conversation so far, so responses stay coherent across turns.
What to do in the meantime
The connection is real and the graph runs — only the context is missing — so anything that doesn't depend on the model remembering still works:
- Pass the previous answer as text. Wire the first node's output into the second node's prompt rather than its Chat History, and write the prompt so it includes what you need carried over. The earlier response then genuinely reaches the model, because it is part of the prompt.
- Keep single-turn steps single-turn. Draft-then-critique pipelines work fine when each step restates what it is critiquing.
Branching conversations
Connect one LLM node's output to the Chat History input of several downstream nodes and each is intended to become an independent branch of the conversation — same shared context, different continuations. Planned uses:
- A/B test prompts against identical context
- Explore multiple dialogue paths side by side
- Compare different models continuing the same conversation
Use cases
Once context is carried, chat history is aimed at:
- Chatbots and assistants that remember the conversation
- Multi-step reasoning — draft, critique, revise as separate turns
- Content pipelines — interview-style generation where each question builds on previous answers
- Education tools — adaptive dialogues that respond to earlier exchanges
Fan-out and parallel does work today, and covers the comparison and A/B patterns above without depending on chat history.
Pinning and Caching
Pin nodes in PixelFlow to cache their output across runs — skip expensive re-execution, iterate faster, and save credits while you refine a workflow.
Fan-Out and Parallel Runs
Run a model once per item with fan-out edges in PixelFlow, then gather results with collect — visual map/reduce for AI: batch images, compare models, scale prompts.