The AI Algorithms Compendium Archives - ChoralSoft

RAG Is a Pipeline, Not a Single Algorithm

alavingi — Wed, 21 Oct 2026 15:42:41 +0000

Retrieval-augmented generation connects a language model to an external collection. Its quality depends on a sequence of design choices, any one of which can become the bottleneck.

Indexing and retrieval

Documents are cleaned, divided into useful chunks, represented for search, and stored with metadata. At query time, lexical, vector, or hybrid retrieval produces candidates. Chunk size, overlap, filters, and embedding choice all affect recall.

Ranking and generation

A reranker can reorder candidates before the prompt is assembled. The generator should distinguish supplied evidence from its own prior patterns and cite the supporting passages when the application requires traceability.

Evaluate each stage

End-to-end answer quality alone does not show where failure occurred. Measure retrieval recall, ranking quality, groundedness, citation accuracy, latency, and cost separately. Also test adversarial documents, stale sources, access controls, and queries with no supported answer.

Continue with the book

Explore the full treatment, worked examples, and exercises in The AI Algorithms Compendium.

The post RAG Is a Pipeline, Not a Single Algorithm appeared first on ChoralSoft.

Q-Learning and Policy Gradients Solve Different Problems

alavingi — Wed, 21 Oct 2026 15:42:31 +0000

← Back to Blog & Articles

Q-learning estimates the value of actions; policy-gradient methods adjust a policy directly. Both optimize long-term reward, but they learn different objects and behave differently in practice.

Q-learning

A Q-function predicts expected return for taking an action in a state and then following a policy. The classic update moves an estimate toward immediate reward plus the discounted best next-state value. It is naturally off-policy and works especially well with manageable discrete actions.

Policy gradients

A parameterized policy outputs action probabilities or continuous-action parameters. The gradient increases the probability of actions associated with higher return. This handles continuous control naturally but often has high-variance updates, so baselines and actor-critic methods are common.

Choose from the environment

Consider action space, sample cost, stability, and whether data comes from the current policy. Reinforcement-learning results are sensitive to reward design and evaluation protocol. Compare multiple seeds and report uncertainty rather than relying on one favorable run.

Continue with the book

Explore the full treatment, worked examples, and exercises in The AI Algorithms Compendium.

The post Q-Learning and Policy Gradients Solve Different Problems appeared first on ChoralSoft.

Why Attention Replaced Recurrence in Many Sequence Models

alavingi — Wed, 21 Oct 2026 15:42:21 +0000

← Back to Blog & Articles

Recurrent networks process sequence positions through a chain of hidden states. Attention creates direct, content-dependent connections between positions, changing both the learning path and the computational trade-offs.

Shorter information paths

In a recurrent model, information from an early token passes through many sequential updates before reaching a distant token. Self-attention can connect those positions in one layer, which helps models learn long-range dependencies.

Parallel training

Tokens in a transformer layer can be processed together once their representations are available. This maps well to modern accelerators. Standard full attention, however, compares every position with every other position and therefore becomes expensive for long contexts.

Architecture is a trade-off

Recurrence remains useful for streaming and bounded-memory settings. Attention is not automatically better for every sequence. The task, context length, latency target, hardware, and data regime determine whether full attention, sparse attention, recurrence, or a hybrid is appropriate.

Continue with the book

Explore the full treatment, worked examples, and exercises in The AI Algorithms Compendium.

The post Why Attention Replaced Recurrence in Many Sequence Models appeared first on ChoralSoft.

Bias and Variance: Diagnosing Model Error

alavingi — Wed, 21 Oct 2026 15:42:11 +0000

← Back to Blog & Articles

Bias and variance describe two different ways a learned model can fail. They are most useful as diagnostic concepts rather than slogans about model complexity.

High bias

A high-bias model cannot capture important structure. Training and validation errors are both poor and often close together. Useful responses include richer features, a more expressive model, or less restrictive regularization.

High variance

A high-variance model fits the training set well but performs much worse on unseen data. More representative data, stronger regularization, augmentation, ensembling, or a simpler hypothesis class may improve generalization.

Use learning curves

Compare training and validation performance as data volume grows. Curves reveal whether collecting more data is likely to help and whether the evaluation split resembles deployment. Always examine subgroup performance and distribution shift; one average error can hide important failures.

Continue with the book

Explore the full treatment, worked examples, and exercises in The AI Algorithms Compendium.

The post Bias and Variance: Diagnosing Model Error appeared first on ChoralSoft.

Bayesian Updating: From Prior Belief to Posterior Evidence

alavingi — Wed, 21 Oct 2026 15:41:55 +0000

← Back to Blog & Articles

Bayesian inference is a disciplined way to revise uncertainty. It combines what was believed before observing data with how likely that data is under competing explanations.

The three ingredients

The prior represents belief before the new observation. The likelihood scores how compatible the observation is with each hypothesis. The posterior is the updated belief after multiplying prior and likelihood and normalizing across alternatives.

A diagnostic-test example

A highly accurate test can still produce a modest probability of disease when the disease is rare. The base rate belongs in the prior; sensitivity and false-positive rate belong in the likelihood. Ignoring the prior is the source of many intuitive mistakes.

Modeling choices matter

Bayesian arithmetic is exact only relative to the stated model. Poor priors, dependent observations treated as independent, or an incomplete set of hypotheses can mislead. Report assumptions and perform sensitivity checks instead of treating one posterior number as timeless truth.

Continue with the book

Explore the full treatment, worked examples, and exercises in The AI Algorithms Compendium.

The post Bayesian Updating: From Prior Belief to Posterior Evidence appeared first on ChoralSoft.

BFS, Uniform-Cost Search, and A*: What Changes?

alavingi — Wed, 21 Oct 2026 15:41:42 +0000

← Back to Blog & Articles

BFS, uniform-cost search, and A* can look nearly identical in code. Their difference is the priority assigned to the next frontier state.

Breadth-first search

BFS expands the shallowest state first. It is optimal when every action has the same cost because depth then equals path cost. A simple queue is sufficient.

Uniform-cost search

When action costs differ, prioritize the lowest accumulated cost g(n). A priority queue replaces the ordinary queue, and a state may need to be reconsidered if a cheaper route is discovered. With nonnegative costs, the first finalized goal has minimum cost.

A* search

A* prioritizes g(n)+h(n), combining cost so far with an estimate of remaining cost. An admissible heuristic never overestimates; a consistent heuristic also respects a triangle-like condition across each edge. Better heuristics can reduce expansions without changing the optimal result.

Continue with the book

Explore the full treatment, worked examples, and exercises in The AI Algorithms Compendium.

The post BFS, Uniform-Cost Search, and A*: What Changes? appeared first on ChoralSoft.