SphereIQ KnowledgeAI™ · Enterprise Memory AI

Production-grade enterprise memory AI your architecture team will trust.

Hybrid retrieval + cross-encoder reranking + permission-aware indexing. 91% answer relevance benchmarked. AWS-native VPC deployment. Sphere has shipped 20+ of these in production.

Stack
AWS BedrockClaude 3.5 SonnetOpenSearch ServerlessHybrid RetrievalBM25 + VectorCohere RerankerTitan EmbeddingsPermission-AwareVPC-Isolated
71%
Vector-only baseline answer relevance
79%
Hybrid search (vector + BM25)
91%
Hybrid + reranker — Sphere default configuration
20+
Enterprise production deployments
// free · no commit
Request an Architecture Review
Sphere’s engineers review your data architecture, source systems, and security requirements. 45 minutes. No commitment.

Sphere responds within 1 business day. Your data and company details are kept strictly confidential.

Architecture

Sphere’s 6-layer enterprise memory stack. Every layer matters.

Most RAG diagrams show three boxes. Sphere’s production architecture has six layers — and skipping any one is why 73% of enterprise AI pilots fail to reach production.

🔌
Ingestion
S3 · Lambda · Glue
→
✂️
Chunking
Semantic · Hierarchical
→
🧮
Embedding
Titan v2 · Cohere
→
🔍
Hybrid Retrieval
Vector + BM25 · OpenSearch
→
📊
Reranking
Cohere Rerank via Bedrock
→
💬
Generation
Claude 3.5 · Nova Pro
sphere_rag_config.yaml — production baseline
# Sphere Enterprise Memory AI — Production Configuration
retrieval:
  strategy: "hybrid"          # vector + BM25
  vector_weight: 0.65
  bm25_weight: 0.35
  candidate_set: 50             # pre-reranking pool
  reranker: "cohere-rerank-v3"
  final_top_k: 8

permissions:
  filter_mode: "pre-retrieval"   # NOT post — critical
  acl_source: "source_system"    # inherited at ingest

index:
  provider: "opensearch-serverless"
  embedding_model: "amazon.titan-embed-text-v2"
  chunk_strategy: "semantic"
  chunk_overlap: 0.15

generation:
  model: "anthropic.claude-3-5-sonnet"
  citation_required: true
  no_answer_handling: "explicit"  # never hallucinate

governance:
  audit_logging: true
  evaluation_cadence: "weekly"
  data_residency: "vpc-only"
Performance Benchmark

The retrieval architecture matters more than the model you choose.

Sphere’s benchmark across 1,200 enterprise queries — same LLM held constant (Claude 3.5 Sonnet), same corpus, different retrieval configurations.

Vector-only (baseline)Dense search, no keyword layer
71%
Hybrid search (vector + BM25)Keyword layer recovers exact-match failures
79%
Hybrid + reranker — Sphere defaultCross-encoder reranking over hybrid candidate set
91%
Hybrid + reranker + metadata filtersDate + doc-type filters for deterministic queries
94%
📌
The practical implication: switching from vector-only to hybrid + reranker adds +20 percentage points of answer relevance without retraining the model, re-embedding the corpus, or changing the LLM. It is a retrieval architecture decision — and it is the first optimization Sphere makes on every new engagement.
Source: Sphere Inc. internal benchmark, Q1 2026. 1,200-query golden dataset. LLM-as-judge evaluation methodology. Corroborated by FloTorch FinanceBench 2026 findings.
Deployment Specification

What Sphere delivers. What you own. What takes 45 days.

What Sphere Builds
  • Source connectors: SharePoint, Confluence, Salesforce, Slack, S3, Google Drive + custom sources
  • Dual-index architecture: OpenSearch Serverless (vector + BM25) inside your AWS VPC
  • Permission inheritance pipeline — ACLs captured at ingest, enforced at retrieval (pre-filter)
  • Hybrid retrieval + cross-encoder reranking via AWS Bedrock
  • Citation-enforcing system prompt with explicit no-answer handling
  • Evaluation suite: RAGAS-based + Sphere extensions (faithfulness, relevance, recall, precision)
  • CloudWatch observability dashboard + weekly evaluation reports
  • Infrastructure as Code (Terraform) for full reproducibility
Architecture Decisions You Control
  • LLM selection: Claude 3.5 Sonnet, Amazon Nova Pro, or on-VPC Llama 3 for air-gapped requirements
  • Vector store: OpenSearch Serverless (AWS-native) or Pinecone / Weaviate if already in stack
  • Embedding model: Amazon Titan v2 (no egress) or Cohere embed-v3
  • Deployment model: managed SaaS, VPC-hosted, or on-premise (regulated industries)
  • Agentic RAG layer: optional for multi-hop reasoning use cases
  • Multimodal ingestion: optional for organizations with significant image/diagram corpora
// free architecture review

45 minutes with a Sphere AI engineer. No commitment.

Sphere reviews your data architecture, source systems, security requirements, and use case — and delivers a concrete recommendation on stack, deployment model, and realistic timeline.

Book My Free Architecture Review

Sphere responds within 1 business day · Available for orgs with 100+ employees

AWS Premier PartnerAnthropic PartnerSnowflake Partner20+ RAG DeploymentsNPS 75