In a RAGFlow chat assistant's system prompt, what does the {knowledge} variable do?

## Where it lives A RAGFlow chat assistant is configured in three parts: the assistant settings (name, opening greeting, empty response, whether to show quotes), the prompt engine, and the model settings. The prompt engine holds the system prompt template plus the retrieval controls that assistant will use — similarity threshold, keyword similarity weight, top N, and optionally a rerank model. The system prompt is a template, not a fixed string, and `{knowledge}` is its reserved slot. ## What substitution actually does On each user turn RAGFlow retrieves from the knowledge bases attached to the assistant, applies the threshold, keeps the top N surviving chunks, and renders them into text. That rendered block replaces `{knowledge}` in the template. The assembled system prompt then goes to the configured model along with the conversation. So the template is where you control everything about how the model is allowed to use the retrieved material: whether it may add outside knowledge, what to do when the block is thin, what format the answer takes, what language to answer in. The retrieval settings decide *which* chunks land in the slot; the prose around the slot decides what the model does with them. ## Other variables `{knowledge}` is the built-in one, but the same configuration screen lets you declare additional named variables. Those are filled per request from the API rather than by retrieval, which is the mechanism for parameterising one assistant across callers — a customer name, a product line, a locale — without cloning the assistant per case. A declared variable that the caller never supplies leaves you with an unfilled template, so keep the declared set and the caller contract in sync. ## The silent failure The single most common mistake is rewriting the system prompt to something more elaborate and dropping the placeholder in the process. Nothing errors. Retrieval still runs and still costs you embedding and search work. The reference list may even still be produced. But the model never sees a document, so it answers from its own parameters — confidently, fluently, and wrongly, in exactly the register that reads as hallucination. Anyone debugging "our RAG is hallucinating" should check the template for the placeholder before touching thresholds or embeddings. ## Interaction with the empty-response setting The assistant settings include an empty response. If retrieval returns nothing that clears the threshold, that fixed text is returned instead of calling the model. Left blank, the model is called anyway with an empty knowledge block and answers unaided. Combining a written empty response with an explicit instruction inside the template — answer only from the material above, and say plainly when it is absent — is what turns "no relevant documents" into a visible, testable outcome rather than a fabricated answer. ## Writing the template well Keep the instructions ahead of the slot and the constraints after it, so the model reads the rules, then the material, then the reminder. Delimit the block visibly. State the refusal rule in one sentence rather than three hedged ones. And keep the template under version control alongside the retrieval settings, because an assistant's behaviour is the pair — a template tuned for six rich chunks behaves differently when someone drops top N to two.