Home / Projects / Evie v3 Knowledge & Retrieval Architecture
Evie v3 Knowledge & Retrieval Architecture
Evie v3 Knowledge & Retrieval Architecture is a project by Akash Renuka Ashok (Co-Founder and CTO). A phased plan and audit that replaces hardcoded AI knowledge and single-purpose retrieval with governed, reviewable infrastructure.
- Role
- Co-Founder and CTO
- Category
- AI Infrastructure
- Stack
- Supabase PostgreSQL, pgvector, Postgres full-text search, Reciprocal rank fusion, gemini-embedding-001, TypeScript
The problem
The shipped assistant's veterinary knowledge existed as 68 hardcoded entries with no provenance, versioning or review workflow, and content changes required a code deployment; retrieval left the knowledge base out entirely and ranked memories by recency rather than meaning.
What Akash owned
Authored the audit and the resulting four-phase architecture plan, specifying the schema, review workflow and retrieval design.
Highlights
- Designed a knowledge schema with source foreign keys, review status, reviewer identity and versioning, backed by a tiered register of permitted sources
- Specified a mandatory veterinary review gate enforced as a TrustOps rule, so retrieval serves only clinically reviewed content
- Designed a unified retrieval service over three corpora using hybrid vector and full-text search merged by reciprocal rank fusion, with confidence-weighted ranking and token-budgeted context assembly
- Specified a retrieval evaluation harness in continuous integration so retrieval quality is measured rather than assumed
Outcome
Produced a concrete, phased path from ungoverned AI knowledge to a source-registered, vet-reviewed knowledge base with measurable retrieval quality.
More work by Akash Renuka Ashok:
Last updated