Glossary

• BM25

  The production-grade evolution of TF-IDF: adds term-frequency saturation and document-length normalization. Still the default ranking function inside Elasticsearch and OpenSearch.

• CNN

  Convolutional neural network. Learns local filters that slide across the input, exploiting spatial structure. The default shape for images before Vision Transformers arrived.

• Constrained decoding

  Restricting which tokens can be sampled at each step, with a grammar or schema, so the output is guaranteed to be well-formed.

• Dense embeddings

  Vector representations from a transformer encoder. The same word gets a different vector depending on context, so semantically related phrases land close together in vector space even with zero shared words.

• Google ADK

  Google's agent framework: tightest integration with Gemini and Vertex AI, native support for the Agent-to-Agent (A2A) protocol.

• Greedy decoding

  Greedy decoding: picking the single highest-probability token at every step, with no search over alternative paths.

• Harness

  The fixed scaffolding that runs a system against a set of cases and reports the result: it owns the inputs, the invocation, and the scoring, so the same run is repeatable. An eval harness is the specific case for a generative system: frozen inputs, a scorer for each property, and a pass rate.

• Hybrid search + reranking

  Run lexical (BM25) and dense embeddings in parallel, merge the rankings (often via reciprocal rank fusion), then optionally rerank the top candidates with a cross-encoder that scores the query and candidate jointly.

• LangGraph

  The most enterprise-adopted stateful agent orchestration framework, from the LangChain team.

• Logits

  A model’s raw, unnormalized scores for each candidate token, before the softmax turns them into a probability distribution.

• Mixture-of-experts

  A layer made of several parallel sub-networks, the experts, plus a router that sends each token to only a few of them. Total parameter count grows while inference cost stays close to that of a much smaller model.

• Pydantic AI / Pydantic Graph

  A typed agent SDK from the Pydantic team. Pydantic Graph specifically is a typed graph-workflow builder (BaseNode, GraphRunContext, End) for modeling routing and state-machine logic.

• Pydantic model

  A Python class (subclass of BaseModel) that defines a data shape via type hints, then validates and parses input against it at runtime instead of trusting an untyped dict.

• RNN

  Recurrent neural network. Processes a sequence one step at a time, carrying a hidden state forward. Stepping through that state sequentially is hard to parallelize, part of why attention displaced it.

• Softmax

  The function that turns a vector of logits into a probability distribution: exponentiate each score and divide by the sum, so the outputs are positive and add to 1.

• State-space model

  State-space model. Processes a sequence through a compressed, continuously updated state, built for the parallel training and long-context efficiency that made Transformers dominant. Mamba is a current example.

• Temperature

  A scalar that rescales a model’s logits before the softmax. Values below 1 sharpen the distribution toward the top tokens; values above 1 flatten it toward the tail.

• TF-IDF

  Term Frequency, Inverse Document Frequency: weights a word by how often it appears in one document versus how rare it is across the whole corpus. Sparse and lexical, no notion of meaning.

• Top-p (nucleus sampling)

  Nucleus sampling. Sorts tokens by probability and samples from the smallest set whose cumulative mass passes a threshold p, so the pool size adapts to the shape of the distribution instead of staying fixed.

• Transformer

  The architecture built on self-attention (block 03). Every token attends to every other token in the window, computed in parallel rather than stepped through sequentially like an RNN.

No terms match that search.