Neocortex RAG
A retrieval platform an AI agent queries as a tool, not through another chat box.
- Client
- Own platform
- Year
- 2026
- Status
- Internal
- My role
- Sole engineer
Problem
Most retrieval products end at a chat window, which means a human has to be the integration. What a team usually wants is for their own agent to reach into the corpus mid-task and keep going.
Architecture
- 01
Ingest asynchronously
Uploads land in S3 and queue on SQS, so a large batch never blocks queries that are already running.
- 02
Embed and index
Bedrock Titan produces the vectors, OpenSearch holds them, and DynamoDB carries the metadata for filtering.
- 03
Expose it twice
A REST console with per-key scoping for people, and an MCP server for models, so any MCP client can use the corpus as a tool.
- 04
Front end
A Next.js console behind Cognito for uploading, browsing and managing API keys.
Impact
Any MCP-capable agent can query the corpus without a bespoke integration
Ingestion scales independently of query traffic
Retrieval is a tool call rather than a person copying answers out of a chat window