v3.4 — live evaluation dashboard just shipped

Sage

The agentic knowledge platform that reads your documents, answers with cited sources, and takes the next action — no human required to close the loop.

Built by EFFOMA, an applied-AI studio. Sage productizes a RAG + multi-agent pipeline reused across 5+ enterprise engagements since 2023.

Try it live See the architecture
81%
recall@5, self-hosted embeddings
BGE-large on a Vietnamese legal corpus, 45ms latency, ~$0 marginal cost
~8 sec
30-page contract analysis
vs ~2 hours of manual legal review, via multi-agent orchestration
~$350/mo
steady-state cost
10,000 docs, ~1,000 queries/day — S3, Lambda, OpenSearch Serverless, Bedrock
70%
inference cost cut
ElastiCache semantic caching on repeated queries, in production
Architecture

Every default is a decision, not a shrug

Six components, six explicit trade-offs. This is what we tell clients when they ask "why this and not that" — no "let the team decide later."

01

Ingestion Pipeline

S3 → EventBridge → Step Functions. The state machine runs Textract (with OCR for scanned pages), chunks by semantic boundary, then embeds before writing vectors to the store.

Trade-off Step Functions over plain Lambda chains buys per-step retry and observability at the cost of added infra complexity. It breaks down at Glue-scale document volume — that's a batch-ETL job, not a state machine.
02

Vector Store

Default: OpenSearch Serverless. Aurora pgvector for small-scale or already-on-Postgres teams. Kendra when enterprise-connector ACLs matter more than retrieval flexibility. S3 Vectors for a cold archival tier.

Trade-off OpenSearch Serverless carries a real ~$200/mo OCU floor regardless of query volume. For light workloads, pgvector is the honest recommendation — not the default.
03

Agent Runtime

Default: Bedrock AgentCore — managed memory and guardrails remove a lot of undifferentiated plumbing. LangGraph on Lambda/ECS when portability across models or infra matters more than managed convenience.

Trade-off This is decided per engagement up front. Deferring it means re-architecting the memory and guardrail layer later — there is no neutral middle ground.
04

Multi-Agent Topology

Supervisor (default): one orchestrator routes to specialist agents — Extract, Compare, Risk, Summary. Hierarchical once you pass ~8 specialists. Swarm rejected for enterprise use.

Trade-off Supervisor costs +1 hop of latency but stays debuggable and auditable. Swarm is faster to prototype but can't answer "why did the system do that" — a dealbreaker for an auditor.
05

Evaluation Pipeline

A 100–200 pair golden dataset drives two metric layers: retrieval (recall@k, MRR/nDCG) and generation (faithfulness via RAGAS / Bedrock Evaluations).

Trade-off Failures are triaged by category before anyone touches a prompt: retrieval miss → fix chunking/embedding, correct retrieval + wrong answer → fix prompt/model, stale answer → fix the sync pipeline. Skip this and you'll re-tune a prompt to fix what was actually a chunking bug.
06

Security & Access Control

Cognito issues group claims, enforced as retrieval-time ACL metadata filters. KMS encryption, VPC endpoints, and CloudTrail wrap the rest.

Trade-off Filtering happens at retrieval time, never after generation — if the model already "saw" data it shouldn't have, blocking the response doesn't erase that from the audit trail. Retrieval-time filtering is the only defensible design in a regulated environment.
Live demo

Try it live

This calls real deployed endpoints — not a mock. Ask a question against the live RAG pipeline, or run the golden-set evaluation suite and watch the metrics come back.

Ask Sage

POST /ask

Evaluation Suite

GET /evaluate

Runs the golden-set retrieval + faithfulness benchmark used to gate every release.

Changelog

Release history

Sage shipped as a named, versioned internal platform starting 2024, with incremental releases through today.

v3.4 Latest July 2026
  • Live evaluation dashboard — recall@5, faithfulness, and per-question triage in real time
  • BGE embedding benchmark tooling: self-hosted BGE-large vs OpenAI ada-002 vs Cohere on a Vietnamese legal corpus
  • Pipeline now reused across 5+ client engagements
v3.0 Early 2026

AgentCore migration

  • Migrated agent runtime to Bedrock AgentCore (managed memory + guardrails)
  • Added supervisor and hierarchical topology support
  • Rolled out multi-tenant deployment for a Vietnamese conglomerate: 5 subsidiaries, 50K documents, 10K queries/day, tenant isolation via Cognito custom claims + per-tenant OpenSearch indices
v2.5 Mid 2025

Evaluation & caching

  • RAGAS integration for automated faithfulness scoring
  • Golden-dataset evaluation harness (100–200 Q&A pairs)
  • ElastiCache semantic caching — ~70% reduction in repeated-query Bedrock inference cost in production
v2.0 Early 2025

Agentic layer

  • Bedrock Agents orchestrator + specialist agents (Extract / Compare / Risk / Summary)
  • Multi-agent contract analysis: 30-page contract in ~8 seconds vs ~2 hours of manual legal review
  • Action-taking added: ticket creation, system lookups
v1.5 Late 2024

Multi-tenant

  • Cognito-based tenant isolation
  • Per-tenant OpenSearch index routing
  • Deployed on a 7-source-system enterprise engagement in Australia (60% scanned/image PDFs, 6-week engagement)
v1.0 Mid 2024

Initial release

  • Single-tenant RAG pipeline, OpenSearch-backed retrieval
  • S3 → Textract → chunk → embed ingestion flow
  • Basic Q&A over internal documents