aicoding.am
Back to cases

Private client system / public summary only

RAG Knowledge Base Assistant

A retrieval-first assistant pattern for searching approved knowledge chunks before answering operational questions.

Problem

Internal knowledge can be scattered across documents, chats and repeated manual answers.

Constraints

Answers must come from approved sources, expose uncertainty when context is missing and avoid private document details in public material.

Solution

Ingest approved sources, chunk content, index embeddings and ground assistant answers in retrieved context.

Role

Architecture, prompt engineering, workflow automation and implementation planning.

Public result

Public summary only: reduced manual lookup through a structured assistant pattern, without client names or unsupported metrics.

Architecture

  1. 01 / Approved source set
  2. 02 / Text normalization
  3. 03 / Chunking strategy
  4. 04 / Embeddings
  5. 05 / PostgreSQL + pgvector
  6. 06 / Retrieval layer
  7. 07 / Strict answer prompt
  8. 08 / Telegram or Web UI
  9. 09 / Logs and review queue

Reusable lessons

  • Retrieval should happen before generation, not after a weak answer.
  • Missing context needs a designed response path.
  • Review logs are part of the system, not an afterthought.
n8nPostgreSQLpgvectorOpenAITelegram BotDocker