EXPLORE / 06

Terms index

Specialized AI and software-development terms with concise explanations

RECORDS

Entity records

72 records · 2/2

  1. 041

    Deterministic Boundary

    The Deterministic Boundary strictly isolates model predictions from hardcoded system logic. While LLMs propose tool intents, authentication, permission checks, cost limits, and Schema validations are unconditionally governed by code to enforce safety invariants.

    Website pending
  2. 042

    Human-in-the-Loop

    Human-in-the-Loop (HITL) pauses automated execution before high-risk actions (e.g., transfers, deployments, mutations) to request explicit human review or parameter edits. It prevents unauthorized model operations and ensures system alignment.

    Website pending
  3. 043

    Checkpointer

    A Checkpointer automatically persists state snapshots after graph node transitions into storage or memory. It enables state resumption, cross-step memory, human-in-the-loop pause/resume, and historical trajectory replay for resilient long-running agent workflows.

    Website pending
  4. 044

    Agent State

    Agent State is a typed schema serving as the single source of truth during an agent's execution lifecycle. It explicitly segregates user inputs, model proposals, retrieved evidence, tool logs, and error codes, preventing reliance on unstructured chat history.

    Website pending
  5. 045

    StateGraph

    StateGraph is a core orchestration structure in LangGraph that models agent workflows as explicit directed graphs. Nodes act as functions updating state, while standard or conditional edges govern transition flows, guaranteeing deterministic control and transparency.

    Website pending
  6. 046

    Router Pattern

    The Router Pattern uses a classifier or LLM router to direct incoming requests to single-purpose deterministic workflows or specialized agents. It offers low latency, high determinism, and straightforward testability for entry-level intent dispatching.

    Website pending
  7. 047

    Plan-and-Execute Pattern

    The Plan-and-Execute Pattern breaks long-horizon goals into a structured sequence of sub-tasks before handing them to an executor, triggering replanning upon failure or new observations. Compared to ReAct, it stabilizes long-chain task execution and mitigates context drift.

    Website pending
  8. 048

    ReAct Loop

    The ReAct Loop (Reasoning and Acting) is an agent pattern where the model alternates between reasoning, selecting tool actions, and observing execution results. It dynamically adjusts steps based on intermediate feedback, suitable for open-ended exploration, but requires step limits and loop guards to prevent execution degradation.

    Website pending
  9. 049

    Ungrounded Generation

    Ungrounded Generation occurs when a language model produces claims that are unsupported by retrieved context passages, leading to hallucinations and factual drift.

    Website pending
  10. 050

    Citation Verification

    Citation Verification validates generated response citations (such as [S1]) against retrieved source passages, checking whether citation IDs exist and strictly support the generated claims.

    Website pending
  11. 051

    Direct Query Search

    Direct Query Search passes unparsed user input directly into vector or keyword search indexes, which frequently fails when queries contain contextual pronouns or multi-intent comparisons.

    Website pending
  12. 052

    Query Rewriting

    Query Rewriting leverages LLMs to transform vague, ambiguous, or conversational user queries into structured, explicit search queries or multiple sub-queries, boosting downstream retrieval recall rates.

    Website pending
  13. 053

    Full Re-indexing

    Full Re-indexing parses, chunks, embeds, and re-indexes the entire document corpus from scratch whenever data changes. While eliminating stale data artifacts, it incurs high compute costs and service downtime.

    Website pending
  14. 054

    Incremental Indexing

    Incremental Indexing uses content hashes to re-embed and update only modified document chunks while purging obsolete records, drastically cutting down compute costs and API overhead during frequent knowledge updates.

    Website pending
  15. 055

    Normalized Discounted Cumulative Gain

    Normalized Discounted Cumulative Gain (NDCG) measures ranking quality by accounting for multi-level relevance scores and logarithmic position decay, providing a comprehensive metric for evaluating search and re-ranking algorithms.

    Website pending
  16. 056

    Mean Reciprocal Rank

    Mean Reciprocal Rank (MRR) evaluates retrieval system performance by measuring how high the first relevant item appears in search results. It is computed as the mean of reciprocal ranks across a set of query evaluation cases.

    Website pending
  17. 057

    Fixed-size Chunking

    Fixed-size Chunking splits text strictly by token or character counts with sliding overlaps. It offers high throughput and easy implementation, but risks breaking logical paragraph structures, code blocks, and table data.

    Website pending
  18. 058

    Parent-Child Chunking

    Parent-Child Chunking index small child chunks to maximize retrieval precision, and maps matched hits back to larger parent document segments for context assembly, balancing precise candidate matching with contextual completeness.

    Website pending
  19. 059

    Score Normalization

    Score Normalization scales vector and BM25 scores into a standardized 0–1 numerical range before applying weighted sums. While preserving relative score margins, it is susceptible to score distribution skew between different search engines.

    Website pending
  20. 060

    Reciprocal Rank Fusion

    Reciprocal Rank Fusion (RRF) combines ranked lists from multiple search algorithms without requiring score normalization across different scales. It scores documents by summing the reciprocal of their ranks across lists, providing robust hybrid search performance.

    Website pending
  21. 061

    Top-K Similarity

    Top-K Similarity directly truncates top candidates purely based on raw vector distance or keyword matching scores. While highly relevant, it frequently retrieves repetitive context passages when indexing overlapping or redundant documentations.

    Website pending
  22. 062

    Maximal Marginal Relevance

    Maximal Marginal Relevance (MMR) is a re-ranking algorithm designed to balance relevance and diversity. It penalizes redundant candidates that are overly similar to already selected documents, avoiding top-k results that suffer from low informational diversity.

    Website pending
  23. 063

    Non-parametric Memory

    Non-parametric Memory refers to explicit knowledge stored outside model parameters, such as in vector databases, document repositories, or knowledge graphs. It enables dynamic retrieval during inference, allowing factual updates without modifying model weights.

    Website pending
  24. 064

    Parametric Memory

    Parametric Memory represents knowledge implicitly stored within neural network weights during pre-training and fine-tuning. While offering rapid access, it remains frozen at training cutoffs, requires computationally expensive retraining to update, and is susceptible to factual hallucination without external context.

    Website pending
  25. 065

    Post-filtering

    Post-filtering retrieves top similarity candidates first and then filters out restricted or mismatched items downstream. This approach risks recall collapse when top candidates are filtered out, while causing unnecessary compute overhead and potential security leak windows.

    Website pending
  26. 066

    Metadata Pre-filtering

    Metadata Pre-filtering applies hard scoping constraints (such as tenant ID, permission ACLs, document version, or update timestamp) prior to executing vector or sparse similarity search. It guarantees unauthorized documents never enter candidate pools, enforcing strict multi-tenant access control.

    Website pending
  27. 067

    Cross-Encoder

    Cross-Encoder feeds a concatenated query-document pair into a single Transformer model to perform joint self-attention across all tokens. It provides significantly higher ranking precision but suffers from high computational latency, making it ideal for second-stage re-ranking rather than initial candidate retrieval.

    Website pending
  28. 068

    Bi-Encoder

    Bi-Encoder processes query and document independently into vector embeddings, allowing document vectors to be pre-indexed for fast sub-second nearest neighbor search across massive corpora, though at the expense of missing fine-grained token-level cross interactions.

    Website pending
  29. 069

    Fine-tuning

    Fine-tuning adapts a pre-trained large language model by updating its neural network weights on task-specific dataset. It improves target output formatting, instruction adherence, and domain-specific stylistic behavior, but incurs high retraining costs and cannot guarantee real-time factual accuracy.

    Website pending
  30. 070

    Retrieval-Augmented Generation

    Retrieval-Augmented Generation (RAG) retrieves relevant external knowledge before passing it into a generative language model to produce grounded responses. It mitigates factual hallucinations stemming from static parametric memory, enables fast knowledge updates, and provides verifiable citation sources.

    Website pending
  31. 071

    Sparse Retrieval

    Sparse Retrieval uses high-dimensional sparse representations and inverted indexes (such as BM25) to score text matching based on term frequency. It is highly effective for exact keyword lookup, alphanumeric IDs, and specialized terminology, but fails to match semantically equivalent queries using different words.

    Website pending
  32. 072

    Dense Retrieval

    Dense Retrieval maps text into low-dimensional dense embedding vectors using neural networks and matches content via vector distance metrics. It excels at semantic similarity, paraphrase matching, and conceptual recall, but may miss precise keyword identifiers or exact alphanumeric terms.

    Website pending