Hi, I am using English knowledge base and multilingual language settings. I am facing an issue in Hindi call. During the call, the word related to query is captured in English but the knowledge base is not retrieved, sometimes. Also in some case, if the knowledge base is retrieved the contents in the chunks are irrelevant. But if the call answered in English, the knowledge base retrieval is done correctly and exact content is retrieved. I can’t find the reason. If there is any way to fix this, it would be helpful.
Hello @mayurigha.s
The knowledge base uses the call transcript to find relevant chunks via vector similarity search. When the call is in Hindi but the KB content is in English, the embeddings may not match well — leading to missed or irrelevant retrievals.
Add Hindi content to your knowledge base — Translate your KB sources into Hindi (or provide bilingual content) so the embeddings align with Hindi transcripts.
Use Markdown format with clear headings and short, focused sections for better chunking.
Add a prompt instruction to guide the agent to rephrase queries in English internally before retrieval, though this has limits.
The core issue is that cross-language retrieval (Hindi transcript → English KB) is inherently less reliable than same-language retrieval.
More details: /build/knowledge-base and /agent/language.
Thank You
Hi, I also tried adding Hindi content to knowledge base, even then the retrieval is not relevant. But in the transcript the query word is captured. In spite of proper query word being captured in transcript, its not able to retrieve the appropriate content in chunks.
Hello @mayurigha.s
Could you please share your agent ID along with the knowledge base (KB) ID and Call ID? This will help our team review your setup more effectively.
Thank you!