Practitioner's Reference Shelf

Resources

Primary papers, production runtimes, and evaluation benchmarks referenced throughout the tutorials and architecture deep dives on largo.dev.

3 Interactive Simulators

Explore all three production AI workbenches in one place: size GPU VRAM and MHA/GQA/Hybrid/MLA KV caches from 2K to 128K context, model two-stage BM25 + dense RRF and cross-encoder reranking latency SLOs, and compare quadratic transcript growth against active-slot agent memory consolidation.

Foundational & Frontier Papers

Reciprocal Rank Fusion Outperforms Condorcet and Individual Rank Learning Methods

Cormack, Clarke & Buettcher, 2009

Introduced rank-based 1 / (k + r) score fusion across lexical and semantic retrieval lists without brittle score-distribution calibration.

Attention Is All You Need

Vaswani et al., 2017

Introduced scaled dot-product and multi-head self-attention, replacing recurrent bottlenecks with parallelizable sequence modeling.

Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks

Reimers & Gurevych, 2019

Reduced pairwise semantic similarity search from hours of cross-encoder inference to milliseconds with bi-encoder vector spaces.

Dense Passage Retrieval for Open-Domain Question Answering

Karpukhin et al., 2020

Demonstrated how dual-encoder training with in-batch and hard BM25 negatives outperforms lexical retrieval alone on open-domain QA.

ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction

Khattab & Zaharia, 2020

Preserves token-level MaxSim matching while precomputing document representations offline, bridging bi-encoder speed and cross-encoder accuracy.

Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

Lewis et al., 2020

Coupled non-parametric dense retrieval indices with sequence-to-sequence generation so grounded facts can update without retraining weights.

LoRA: Low-Rank Adaptation of Large Language Models

Hu et al., 2021

Freezes pre-trained model weights and injects trainable rank-decomposition matrices into linear layers, enabling compact multi-tenant serving adapters.

FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning

Dao, 2023

IO-aware exact attention tiling across GPU SRAM and HBM that makes long-context training and serving practical at high FLOP utilization.

Mamba: Linear-Time Sequence Modeling with Selective State Spaces

Gu & Dao, 2023

Input-dependent state space selection with hardware-aware parallel scan, achieving linear O(N) scaling on long sequences and telemetry streams.

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

DeepSeek-AI, 2025

Showed how Group Relative Policy Optimization (GRPO) against verifiable rewards elicits multi-step reasoning and self-verification without dense process supervision.

mHC: Manifold-Constrained Hyper-Connections (arXiv:2512.24880)

DeepSeek-AI, 2025

Projects multi-stream residual mixing matrices onto the Birkhoff polytope via Sinkhorn-Knopp iterations, bounding activation gain across deep MoE training.

Tools & Runtime Stack

PyTorch & torch.compile

Runtime

Core tensor runtime and TorchInductor graph compiler for training, quantization, and sub-millisecond kernel fusion.

vLLM

Serving

High-throughput LLM serving engine with PagedAttention, continuous batching, speculative decoding, and prefix caching.

Pydantic

Contracts

Schema-first validation boundary for typed tool contracts, structured model outputs, and runtime configuration.

sentence-transformers

Embeddings

Reference library for training, fine-tuning, and benchmarking bi-encoder embeddings and cross-encoder rerankers.

FAISS

Vector Search

C++ and GPU library for HNSW, IVF, and product-quantized vector indexing when testing recall-latency trade-offs.

Hugging Face Transformers

Models

Model definitions, tokenizers, and safetensors weight loading across open-weight language, vision, and multimodal architectures.

Hugging Face PEFT & TRL

Fine-Tuning

Parameter-efficient LoRA/QLoRA adapter injection and supervised post-training pipelines for open-weight Transformer and MoE backbones.

ONNX Runtime

Inference

Cross-platform inference execution provider for CPU/GPU graph optimization, INT8 quantization, and portable deployment.

Model Context Protocol (MCP)

Agents

Open JSON-RPC specification standardizing how agent runtimes discover and invoke typed tools, resources, and prompts.

Google Agent Development Kit (ADK)

Agents

Code-first Python and Java framework for typed agent workflows, session state scopes, and managed Vertex AI Memory Bank integration.

LangGraph

Agents

Explicit state-machine runtime for cyclic and parallel agent graphs, checkpointed execution, and conditional edge routing.

Evaluation & Retrieval Benchmarks