AI engineering case study
500,000 technical documents. Engineers spending hours on manual search. We built a hybrid retrieval system that returns accurate answers in under 200ms at 35% lower cost than the initial cloud spec.
500k+
Technical documents indexed
sub-200ms
Query latency in production
35%
Lower cloud cost vs. initial spec
99.9%
Uptime on EKS over 6 months
The problem
Rivian's engineering teams produced over 500,000 technical documents: service manuals, supplier specs, internal wikis, compliance records. Engineers needed answers fast. What they got instead was a keyword search that surfaced documents, not answers.
The time cost was measurable. Senior engineers were spending 45 to 90 minutes per query sifting results, cross-referencing documents, and verifying information manually. At scale, that is an enormous productivity drag. The ask was simple: make the knowledge findable.
Architecture
What we built
We designed a hybrid retrieval pipeline from scratch. Dense semantic search via Weaviate caught conceptually related documents even when the engineer didn't know the exact terminology. Sparse BM25 search caught precise keyword matches for part numbers, model codes, and spec values. The two results were fused using Reciprocal Rank Fusion, then filtered through a cross-encoder re-ranker to surface the top context before hitting the LLM.
Ingestion was built as an async delta-sync pipeline: documents enter a queue, get chunked with context-aware overlap, embedded in batches, and written to Weaviate. New and updated documents flow through without full re-indexing. The entire 500k document corpus was indexed in under 4 hours.
We instrumented every layer with OpenTelemetry, feeding latency, retrieval quality, and LLM cost into Grafana dashboards built specifically for AI workloads. Ragas evaluation ran in CI on a curated question set, so any regression in retrieval quality blocked the deploy.
Tech stack
Retrieval
Weaviate (dense vector search), BM25 (sparse keyword), hybrid fusion with RRF
Embedding
OpenAI text-embedding-3-large, batched ingestion pipeline with delta sync
Re-ranking
Cross-encoder re-ranker over top-50 candidates before final context assembly
Orchestration
LangGraph for multi-step retrieval chains, query expansion, and fallback routing
Infrastructure
AWS EKS with Karpenter autoscaling, Terraform modules, Grafana for AI-specific observability
Evaluation
Ragas for RAG faithfulness and answer relevance scoring, continuous eval in CI
We scope every AI project before we quote. Bring your use case and we'll tell you exactly what it takes to ship it properly.