-
What Is Inference?
-
The Information Selection Problem
-
What Context Means
-
Beyond the Chat Box
-
What Remembering Means
-
From Training to Execution
-
Documents, Identity, and Provenance
-
The Measurement Instrument
-
Never Stand in Front of the Steamroller
-
The Measurement Instrument
-
Tokens, Embeddings and Probability
-
Building an Inverted Index
-
The Context You Didn't Type
-
If There's Any Doubt, It's Deterministic
-
The RAG Baseline
-
The Transformer as a Computation
-
BM25: Ranking with Terms
-
The Context Window Is a Budget
-
From Retrieval to Persistent Understanding
-
Scrum Built the Training Set
-
Autoregressive Generation
-
The First Search Baseline
-
More Is Not Better
-
Meat Proxy
-
Pathways Through Memory
-
Building Our First Inference Engine
-
Parsing, Chunking, and Retrieval Units
-
Order Changes Meaning
-
The Memory Nexus
-
The Price of Intelligence
-
Matrix Multiplication and GPU Execution
-
Dense Retrieval without Re-teaching Embeddings
-
Not All Tokens Are Equal
-
Why
-
Intelligence in the Wrong Direction
-
Attention and Its Computational Cost
-
Exact Search, ANN, and Index Choice
-
Make the Window Bigger
-
Is It Still True?
-
Where Are the Finished Projects?
-
Prefill and Decode
-
Filters, Scope, and Freshness
-
Reuse Before Recompute
-
One Runtime, Many Windows
-
What Did We Leave Unfinished?
-
Key-Value Caching
-
Semantic Failure Analysis and Hard Negatives
-
Remove What No Longer Matters
-
A Revolver, Not a Foundation
-
What Matters Right Now?
-
Memory Bandwidth and Compute Bottlenecks
-
Candidate Generation as a Pipeline
-
Compression Is Loss
-
Consequences Nobody Wrote Down
-
The Smallest Useful Model Call
-
Quantization and Numerical Precision
-
Hybrid Search and Reciprocal Rank Fusion
-
Retrieval From First Principles
-
Let Old Context Fade
-
Does Better Context Change Behaviour?
-
One Operation, Several Model APIs
-
FlashAttention and Optimized Kernels
-
Inference From First Principles
-
Learned Sparse and Multi-Vector Retrieval
-
Externalize the Working Set
-
Token Counts Don't Add Up
-
When the Frame Is Wrong
-
PagedAttention
-
Rerankers and Cross-Encoders
-
Bring Back Only What You Need
-
A Successful Call Is Not Finished Work
-
Context Is a Bottleneck
-
Continuous Batching
-
Diversity, Coverage, and Evidence Budgets
-
The Agent Writes Its Own Context
-
What Did the Model Actually See?
-
What Should Memory Keep?
-
Prefix Caching
-
Understanding Queries and Routing
-
Memory Is a Context Source
-
Learning From Experience
-
Restart Is Not Resume
-
Speculative Decoding
-
Query Rewriting and Expansion
-
Tools Produce Context
-
Can Memory Be Trusted?
-
Preserve Before You Interpret
-
Compilation and Execution Graphs
-
Decomposition and Multi-Hop Retrieval
-
When Memory Becomes Documentation
-
Representation Is Part of Context
-
Claims, Evidence, and Decisions
-
Mixture-of-Experts Inference
-
Hierarchical and Parent-Child Retrieval
-
Who Gets to Be Right?
-
The Agent Said Done. Did Anything Change?
-
The Remembering System
-
Expert Routing and Load Balancing
-
Graph and Structured Retrieval
-
Context Goes Stale
-
Appendix — Building a Remembering Agent
-
Capability Is Not Authority
-
Multi-Head and Multi-Latent Attention
-
Bounded Adaptive Retrieval: When to Stop
-
Edition History
-
Keep Context in the Right World
-
The Agent Cannot Grade Its Own Homework
-
Sliding-Window Attention
-
Relevance Judgments and Test Collections
-
Assemble for the Task
-
Retries Are Side Effects Too
-
Long-Context Inference
-
Metrics that Answer Different Questions
-
The Context Compiler
-
Blind Before You Compare
-
Model Architecture Versus Runtime Architecture
-
Controlled Experiments and Regression Gates
-
What a Compilation Proves
-
The Models Were Different. Their Mistakes Weren't
-
From Inference Loop to API
-
Latency, Cost, and Operational Reliability
-
Did It Reach the Model?
-
When Your Benchmark Is Too Easy
-
Request Scheduling and Admission Control
-
Does Better Retrieval Improve the Task?
-
Did the Context Help?
-
Discovery Is Not Promotion
-
Concurrent Requests and Streaming
-
The Reference Retrieval Contract
-
The Complete Context System
-
Replicate Before You Believe
-
Distributed Inference
-
Defending the Retrieval Boundary
-
What Should Happen Next?
-
Prefill/Decode Disaggregation
-
Improve and Writer: A Real Retrieval Workflow
-
One Process, End to End
-
Reliability, Cancellation and Recovery
-
What We Can Now Prove
-
Your Applied AI
-
Inference Protocols and Harmony
-
When the Process Learns
-
Claude, Caching and API-Level Optimization
-
Reasoning Models and Inference-Time Compute
-
Tool Calling and Agentic Inference
-
Structured Outputs and Constrained Decoding
-
Latency, Throughput and Time to First Token
-
Cost per Million Tokens
-
GPU Utilization and Capacity Planning
-
Correctness, Determinism and Reproducibility
-
Designing Meaningful Inference Benchmarks
-
Designing the Inference Runtime
-
Implementing Scheduling and Caching
-
Adding Multiple Model Backends
-
Benchmarking Against Established Engines
-
What We Learned: Inference From First Principles