Portfolio Ops

RAG console

Ask questions about this portfolio, inspect what the retrieval system actually did, and browse the eval suite.

Backend

Live Bedrock RAG · what this proves

A hand-rolled bounded agentic loop straight against AWS Bedrock (Nova Pro): the model decides per turn whether to retrieve again from Qdrant hybrid search, capped at three iterations, with a forced-grounding and relevance-floor guardrail before any answer ships.

Honesty note: Bedrock already remembers. Every query here has kept a rolling 10-turn window server-side and rewritten conversational follow-ups against it — the Chat / Single question toggle below makes that real, existing behavior visible and lets you opt out of it per question, rather than adding memory that wasn't there before. Chat transcripts are kept for a year (see the retention note in Inspect); single-question turns are not recorded at all.

Architecture

4 layers

Ask console

Bedrock Nova Pro loop

Cited answer

Read the write-up →Full architecture →Why chat mode works this way (ADR-015) →

The 5-turn cap scopes a conversation; it isn't an abuse control — starting a new chat gives you another 5, by design. What actually bounds spend is the shared daily Bedrock counter and the account budget behind it. The per-IP sub-cap is fairness, not security: rotating IPs defeats it and it fails open on a storage error, so one determined visitor could still use up a day. EKS runs are bounded separately, by their own approval and teardown lifecycle, and don't draw on this counter.

After asking, switch to the Inspect tab to see retrieval diagnostics for this query.

AboutEMpathWritingProductionConnect