OyeChatsBetter answers start before the answer.
Grounded conversations, more dependable retrieval, and a shorter path from question to useful answer.
- 01Question
- 02Retrieve context
- 03Build prompt
- 04Stream answer
Better answers start before the answer.
OyeChats is a conversational product that answers from a website’s content, qualifies leads, and supports human handoff. A useful conversation depends on the retrieval and execution paths underneath it.
The problem to work through
Useful knowledge-base content was being excluded by a relevance gate. Duplicated retrieval paths and oversized assembled prompts also made the behavior harder to reason about.
Retrieval diagnosis
I traced the relevance gate that excluded useful content, following the path from retrieved material to the answer rather than treating every refusal as a prompt problem.
A simpler execution path
I consolidated duplicated RAG pipelines and prompt assembly, and worked on streaming and durable qualification jobs so the conversation and its follow-up work could be handled more deliberately.
Evaluation-driven corrections
Later changes addressed regressions found in question evaluations, including company-context answers and long whitespace runs that could block a worker.
15,233 to 8,120 tokens
One assembled prompt, as documented in the remediation pull request. This is a prompt-size comparison, not a latency or answer-quality benchmark.
What I take from it
Retrieval gates, prompt assembly, and background execution all shape the answer a person sees. Diagnose that whole path before adding another instruction.