Resources — high-trust sources for AI Infrastructure

Ground every non-obvious claim in these; cite inline (<sup> → #sources). Populated before teaching. Prefer primary sources (papers, vendor engineering docs) over secondary commentary. This track is the infra/ops deep dive — serving, gateways, scaling, cost — so the sources skew toward systems papers and product docs, not modeling papers.

Primary specs / docs

• pgvector — documentation (github.com/pgvector/pgvector). Vector search inside Postgres: exact + HNSW/IVFFlat indexes. The “do I even need a dedicated vector DB?” baseline — relevant when retrieval volume fits in your existing OLTP store.
• Pinecone — documentation (docs.pinecone.io). Managed serverless vector DB; useful for the build-vs-buy and operational-model (sharding, replication, metadata filtering) framing.
• Weaviate — documentation (weaviate.io/developers/weaviate). Open-source vector DB with hybrid search and built-in reranking modules; good reference for the hybrid + rerank pipeline.
• vLLM — documentation (docs.vllm.ai). Production inference server: PagedAttention, continuous batching, KV-cache management, prefix caching. The reference implementation for the serving lesson.
• Anthropic — prompt caching (docs.anthropic.com). KV/prefix caching at the API layer: cache the stable prefix (system + tools), pay reduced rates on cache hits. The cost lever for L5.
• OpenAI — embeddings & API docs (platform.openai.com/docs). Embedding model dims/metrics and the hosted-embeddings cost/latency baseline for the embeddings lesson.

Papers & deeper reading

• Johnson, Douze & Jégou, 2017 — “Billion-scale similarity search with GPUs” (arXiv:1702.08734; FAISS). Why exact k-NN does not scale and how IVF + product quantization trade recall for memory and speed at billion-vector scale.
• Malkov & Yashunin, 2016 — “Efficient and robust approximate nearest neighbor search using Hierarchical Navigable Small World graphs” (arXiv:1603.09320; HNSW). The graph-based ANN index most vector DBs default to; the recall ↔ latency ↔ memory knobs (ef, M).
• Reimers & Gurevych, 2019 — “Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks” (arXiv:1908.10084). Why you pre-compute embeddings once (bi-encoder) instead of scoring every pair at query time — the architecture that makes vector search feasible.
• Khattab & Zaharia, 2020 — “ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT” (arXiv:2004.12832). Late-interaction retrieval: the middle ground between cheap bi-encoders and expensive cross-encoder rerankers; storage cost trade-off.
• Kwon et al., 2023 — “Efficient Memory Management for Large Language Model Serving with PagedAttention” (arXiv:2309.06180; vLLM, SOSP ‘23). KV cache as the serving bottleneck; OS-paging applied to attention. The spine of the inference-serving lesson.

Notes on trust

• Vendor docs describe their product’s current behaviour; treat product specifics (default index type, exact cache discount, supported metrics) as version-dependent and verify against current docs before teaching as fact.
• Throughput/latency/cost numbers in serving papers are hardware- and model-specific. Teach the shape of the trade-off; quote concrete numbers only as illustrative “at time of writing.”