SaaS Content Platform Cuts Vendor Lock-in and Scales to 100M+ Documents by Migrating from Pinecone/OpenAI to AWS OpenSearch and Claude
At a glance
A content-intelligence platform serving marketing and sales teams was running a dual-vendor search stack — Pinecone for vector storage, OpenAI for embeddings—that was reaching its limits as query volumes grew and agentic AI use cases emerged. Rising SaaS costs, manual infrastructure operations, and an inability to implement hybrid search were creating a ceiling on both growth and product capability.
New Math Data migrated the platform to a fully AWS-native search stack: production-grade OpenSearch 2.19 with Amazon Bedrock and Claude embeddings, hybrid search, Terraform automation, and hardened multi-AZ security—delivered through AWS’s Migration Acceleration Program.
The result: a scalable, compliance-ready foundation handling 100M+ documents and hundreds of thousands of burst queries, purpose-built for the client’s next chapter as an agentic content-intelligence platform.
Industry
Use Case
- Migration and Implementation (Pinecone/OpenAI to AWS)
- Natural Language Processing
- Intelligent Document Processing
- Knowledge Management
Solution implemented
- Three-node OpenSearch 2.19 cluster across three Availability Zones on r6g.large.search instances with 300 GB encrypted EBS per node
- Hybrid search combining BM25 lexical retrieval with Amazon Bedrock Claude vector embeddings, replacing OpenAI
- Shard and rollover strategies (~10 GB shards, 100 GB/7-day rollover) optimized for tens of millions of documents at 900 GB scale
- Full Terraform IaC with S3 and DynamoDB state management eliminating infrastructure drift
- Three-tier network model (Public API / Protected App / Private Data) with CloudFront + WAF, ECS Fargate, and private data tier
- Centralized CloudWatch logging and SNS alerting across all services
The value equation
- Eliminated dual vendor lock-in (Pinecone + OpenAI), consolidating to a single AWS-native stack and removing the scaling ceiling
- Unlocked hybrid search (keyword + vector semantics) enabling agentic and RAG-based workflows previously impossible on Pinecone
- Reduced engineering maintenance burden through full Terraform automation, eliminating manual infrastructure operations
- 99.9%+ uptime architecture across three Availability Zones, replacing a system with known latency degradation under load
- Established a future-ready foundation for AI-driven personalization and MCP-based enterprise agent integrations
Company Snapshot
A US-based content-intelligence SaaS platform helping marketing and sales professionals discover, curate, and share relevant content to build trust and drive revenue.
Location
United States
Customer Situation
The client had built a semantic search index on Pinecone with OpenAI embeddings, but as customer base and query volumes grew, the architecture showed strain. Engineering time was consumed by manual operations rather than product development, and the platform lacked hybrid search capabilities needed to power emerging agentic workflows.
The urgency was strategic: the client was positioning its search index as a RAG-based backend for enterprise agent platforms. Supporting hundreds of thousands of burst queries, enabling natural language search alongside structured filtering, and establishing a credible cost-per-query pricing model for enterprise customers all required a fundamentally different infrastructure foundation. Dual dependency on Pinecone and OpenAI created vendor risk, unpredictable SaaS costs, and integration complexity that constrained growth.
NMD Solution
New Math Data assessed the existing Pinecone and OpenAI stack through an AWS MAP engagement, documenting the migration strategy across the 7Rs framework and designing the target AWS-native architecture. The assessment confirmed that migrating to OpenSearch with Claude embeddings via Bedrock would eliminate both vendor dependencies while delivering capabilities the prior stack could not support.
NMD re-embedded tens of millions of documents through parallel batch processing using Amazon Bedrock Claude, exported all vectors from Pinecone, and loaded them into a newly designed OpenSearch 2.19 cluster optimized for 900 GB of data and burst query patterns. A phased dual-read validation approach enabled zero-downtime cutover, with Pinecone retained in read-only mode as a fallback until full production validation.
Beyond the migration, NMD delivered full Terraform automation, a three-tier network security model, automated credential rotation, and centralized observability — establishing the compliance and reliability posture needed for regulated-industry enterprise customers.
What We Delivered
New Math Data replaced the client’s dual-vendor Pinecone/OpenAI stack with a production-ready AWS-native search platform supporting tens of millions of documents at 900 GB scale, burst query volumes in the hundreds of thousands, and hybrid search combining semantic and keyword retrieval—capabilities that did not exist on the prior architecture.
The migration established the infrastructure foundation for the client’s agentic content-intelligence product roadmap, with a follow-on engagement now underway to build a cost-per-query modeling tool enabling the client to price and pre-sell agentic search capabilities to enterprise customers.
Ready to Transform Your SaaS Business?
See how New Math Data can transform your SaaS business with AWS-powered innovation.