DoorDash Cut Chatbot Hallucinations 90% With a Two-Tier Guardrail, Not a Better Model

DoorDash cut hallucinations in their Dasher support chatbot by 90% and severe compliance issues by 99%. The fix was not a better model. Here's how their team built the quality layer around it: 1. Why the Old System Capped Out The old flow relied on pre-built resolution paths, so only a small subset of issues could be resolved. The knowledge base behind it had three problems: the right article was hard to find, the useful part inside took time to locate, and every article was English while many Dashers prefer another language. 2. Retrieval Starts by Condensing the Conversation An issue is spread across several messages, so the system first condenses the conversation into the core problem. That summary searches historical data for the top N similar cases already resolved with KB articles, and each match maps to an article that goes into the prompt template. Summary quality gates everything downstream. 3. Two-Tier Guardrail, Cheap Check First They tried a sophisticated guardrail model first. Response times and token usage made it prohibitively expensive. Shallow check → sliding-window semantic similarity between response and article segments. A close match is less likely to be a hallucination. LLM evaluator → runs only when the shallow check flags. Sees the response, the KB articles and the conversation history, then judges groundedness, coherence, compliance. Every test must pass before a Dasher sees the response. 4. What LLM Judge Measures They manually reviewed thousands of transcripts, then categorised quality into five aspects: retrieval correctness, response accuracy, grammar and language, coherence to context, relevance to the request. LLM Judge is prompted with open-ended questions. High-frequency answers become the rules used for monitoring. A human team still reviews random samples. Key Takeaways 1. The guardrail is an architecture decision, not a prompt. Cheap check first, expensive check only on suspicion. 2. Latency is the cost. Generate, guard, sometimes retry. They route to human agents rather than wear it. 3. They found the failure categories by reading transcripts, then automated what they found. Reliability here came from everything wrapped around the model, not the model itself.

Confirmed accurate:

• The 90% hallucination reduction and 99% severe compliance issue reduction figures are DoorDash's own reported numbers.This guardrail system has successfully reduced overall hallucinations by 90% and cut down potentially severe compliance issues by 99% DoorDash
• Old system relied on flow-based/decision-tree resolution paths, replaced by RAG — accurate.
• Two-tier guardrail: cheap shallow check first, expensive LLM evaluator only on suspicion — accurate. Initially, we tested a more sophisticated guardrail model but increased response times and heavy usage of model tokens made it prohibitively expensive. Instead, we adopted a two-tier approach: a cost-effective shallow check developed in-house followed by an LLM-based evaluator DoorDash
• Shallow check = semantic similarity between response and KB article segments — accurate. The first quality check layer is a semantic similarity comparison between the response and KB article segments. DoorDash
• Five quality dimensions from manual transcript review — accurate. We manually reviewed thousands of chat transcripts between the LLM and Dashers to develop an iteration pipeline for monitoring LLM quality. Our review allowed us to categorize LLM chatbot quality aspects into five areas DoorDash
• Guardrail checks groundedness, coherence, and compliance — accurate. It checks the grounding of RAG information to prevent hallucinations, maintains response coherence with previous conversations, and filters out responses that violate company policies. DoorDash
• caveats:

• The 90% figure quoted in some places actually comes from offline simulation testing, not solely live production — worth noting since the post presents it as a clean production outcome. Ultimately, we were able to reduce hallucinations in our simulations by 90%; this result carried over into production. DoorDash
• Independent reviewers flag that DoorDash's post doesn't disclose baseline rates or exact measurement methodology — so the percentages, while real, lack context for how dramatic the underlying change actually was.
• The post omits DoorDash's third layer — a simulation flywheel testing 200+ synthetic conversations pre-deploy — which is a meaningful part of the actual system architecture, so the summary is solid but incomplete rather than wrong.