EERRAA // KNOWLEDGE
It answers from your docs, not the internet.
// upload your product manuals, help center, and policies. the agent retrieves the right passage every time, and cites what it read.
vector + full-text, weighted fusion
cross-encoder, 0.30 threshold
best chunks injected per answer
Arabic / English / mixed
HOW RETRIEVAL WORKS
Upload PDF, TXT, or MD. Files are chunked and embedded with BAAI/bge-m3 (1024-d), which handles Arabic, English, and mixed text in the same vector space.
Every question runs two searches at once: vector cosine top-K over pgvector, and Postgres full-text. The two result sets are fused with a weighted score.
A cross-encoder reranks the fused candidates, then a 0.30 relevance threshold trims the weak ones. The top four chunks are injected into the answer.
TWO AUDIENCES
A shared pool for everyone, and a private one per user.
The "project" audience. Docs everyone can draw on: your product manuals, help center, policies. Upload once, every user benefits.
The "user:<id>" audience. Each signed-in user uploads their own documents, retrieved only for them. Their files never leak into anyone else’s answers.
One embedding space for every language.
Ask in Arabic about an English doc, and it still finds the passage.
Feed it your docs.
// knowledge base is in the free tier. upload a file and ask a question.