Appendix C · Glossary

One or two lines per term. Chapter numbers point to where the term is used in earnest.

Adaptive RAG: Routing a question to no-retrieval, single-step or multi-step retrieval based on a predicted complexity. Ch 15.

Agent: An LLM in a loop that chooses actions (tool calls) based on observations until a stop condition. Ch 14–16.

Agentic RAG: RAG where retrieval is a tool the model decides to call, grade and repeat, rather than a fixed step. Ch 15.

ANN (approximate nearest neighbour): Search that trades a little recall for speed by not comparing against every vector (HNSW, IVF). Ch 4.

Answer correctness: Agreement of the generated answer with a reference answer. App B.

Answer relevance: Whether the answer addresses the question asked, regardless of truth. App B.

API key: A secret token identifying a caller to a service; hashed at rest, rotatable. Ch 17.

Atomic claim: A single verifiable statement extracted from an answer for faithfulness checking. Ch 10, 11.

Backoff (exponential): Retrying after waits that double each time, with jitter, to survive rate limits. Ch 16.

Bi-encoder: Two independent encoders (query, document) whose outputs are compared with cosine/dot; precomputable. Ch 2, 9.

Binary quantisation: Storing each vector dimension as one bit; ~32× smaller; needs rescoring. Ch 18.

BM25: Sparse lexical ranking function: IDF × saturated term frequency × length normalisation. Ch 8.

Bootstrap confidence interval: Resampling the per-question scores with replacement to get an interval around a metric; the honest way to report a small eval set. Ch 21.

Bounding box (bbox): The rectangle giving a text block's or figure's position on a PDF page; the raw material for reconstructing reading order. Ch 28.

Checkpointer: LangGraph component that persists graph state per thread after every step (InMemorySaver, Postgres). Ch 14.

Chunk: A piece of a document small enough to embed and retrieve as a unit. Ch 5.

Chunk overlap: Characters/tokens shared between consecutive chunks so facts at boundaries are not split. Ch 5.

Citation: A reference from an answer sentence to the chunk that supports it, usually by numbered id. Ch 11.

Cohen's kappa: Agreement between two raters (your LLM judge and a human) corrected for chance agreement; how you calibrate a judge. Ch 26.

ColBERT / late interaction: Retrieval that keeps one vector per token and scores with MaxSim. Ch 19.

Collection: Qdrant's unit of storage: a set of points with the same vector configuration. Ch 4.

ColPali: Late-interaction retrieval over page images using a vision-language model; no OCR. Ch 19.

CombSUM / CombMNZ: Score-fusion baselines: sum normalised scores across retrievers; CombMNZ additionally multiplies by how many retrievers returned the document. Ch 23.

Conditional edge: A LangGraph edge whose target is chosen by a function of the state. Ch 14.

Context engineering: Deciding what goes into the model's context at each step (instructions, retrieved text, tool results, memory) under a token budget. Ch 19.

Context precision / recall: Judged usefulness of retrieved chunks / coverage of the reference answer by the context. App B.

Context window: Maximum tokens a model accepts in one call. Ch 1.

Contextual retrieval: Prepending an LLM-written situating sentence to each chunk before indexing (Anthropic, 2024). Ch 9.

Cosine similarity: Dot product of two vectors divided by their magnitudes; 1 = same direction. Ch 2.

CRAG (Corrective RAG): Grade retrieval as correct/ambiguous/incorrect and take a corrective action (refine, web search). Ch 15.

Cross-encoder: A model that reads query and document together and outputs a relevance score; precise, slow; used for reranking. Ch 9.

Dead-letter queue (DLQ): Where a work item goes after exhausting its retries, so one poison document cannot block a pipeline. Ch 30.

Dense retrieval: Retrieval by similarity of learned embeddings. Ch 2–4.

Dimensions: Length of an embedding vector (1,536 for text-embedding-3-small). Ch 2.

Distance metric: Cosine, dot product or Euclidean; chosen at collection creation. Ch 4.

Document loader: LangChain component that reads a source (PDF, directory, web) into Document objects. Ch 6.

Dot product: Σ aᵢbᵢ; equals cosine for unit vectors. Ch 2.

ef_construction: HNSW build-time beam width: how many candidates are considered when inserting a node. Higher means a better graph and slower indexing; fixed once built. Ch 22.

ef_search (hnsw_ef): HNSW query-time beam width: how much of the graph a search explores. The one HNSW knob you can change without rebuilding; raises recall and latency together. Ch 22.

Embedding: A vector representation of text such that semantic similarity ≈ geometric closeness. Ch 2.

Ensemble / hybrid retrieval: Combining dense and sparse results, usually by RRF. Ch 8.

Entailment (NLI): Deciding whether a claim is entailed, contradicted or simply not mentioned by a passage; a stricter basis for faithfulness than a yes/no 'supported' judgement. Ch 26.

Eval (evaluation) suite: Golden questions plus metrics run automatically after every change. Ch 10.

Faithfulness / groundedness: Fraction of answer claims supported by the retrieved context. Ch 10, 11.

FastEmbed: Qdrant's lightweight local embedding library; used here for sparse BM25 vectors. Ch 8.

Few-shot prompting: Including worked examples in the prompt to teach format and behaviour. Ch 20.

Filter (payload filter): Qdrant condition on metadata applied during vector search (must, should, must_not). Ch 4.

Fine-tuning: Updating model weights on task data; good for style/format, not for injecting changing facts. Ch 1.

Fusion: Merging ranked lists from several retrievers (RRF, DBSF). Ch 8.

Golden dataset: Curated questions with reference answers and source ids used for evaluation. Ch 10.

GraphRAG: RAG over an LLM-extracted knowledge graph with community summaries for global questions. Ch 19.

Grounding: Constraining the model to answer only from provided context. Ch 11.

Guardrail: A check or limit around the model: PII filters, call limits, injection defences, human approval. Ch 16.

Hallucination: Output not supported by the context or the facts. Ch 11.

Hard negative: A training example that looks relevant but is not; sharpens embedding models. App A (DPR).

Hit rate@k: Whether any relevant chunk is in the top k. App B.

HNSW: Hierarchical Navigable Small World graph; the ANN index in Qdrant. Ch 4, 18.

Human-in-the-loop: Pausing an agent for approval before a consequential action (interrupt). Ch 16.

HyDE: Embedding a hypothetical LLM-written answer instead of the question. Ch 9.

Idempotent upsert: Writing with stable ids so re-running ingestion overwrites instead of duplicating. Ch 7.

IDF (inverse document frequency): log(N/df); weights rare terms higher in BM25. Ch 8.

In-memory vector store: LangChain's InMemoryVectorStore; fine for a first RAG, not for production. Ch 3.

Indirect prompt injection: Instructions hidden in retrieved documents or tool outputs that the model may follow. Ch 19.

Information bottleneck (bi-encoder): A bi-encoder must compress a document into one vector before seeing the query, so it encodes general topicality rather than the specific condition asked about. The reason rerankers exist. Ch 24.

Ingestion: The offline pipeline: load → parse → chunk → embed → store. Ch 6, 7.

IVF: Inverted file index: cluster vectors, search a few clusters. Ch 18.

JWT: Signed JSON token carrying identity and expiry, sent as a bearer token. Ch 17.

k (top-k): Number of chunks retrieved (or returned after reranking). Ch 3.

LangChain: Framework providing models, messages, tools, documents, splitters, vector-store integrations and create_agent. Throughout.

LangGraph: Library for building stateful graphs (nodes, edges, checkpoints) that run agents. Ch 14–16.

LangSmith: Tracing and evaluation platform for LangChain/LangGraph runs. Ch 13.

Late chunking: Embedding the whole document first, then splitting the token embeddings into chunks. Ch 19.

LazyGraphRAG: GraphRAG variant that defers LLM summarisation to query time; indexing cost like vector RAG. Ch 19.

Lease (work item): A time-limited claim on a queued item so a crashed worker's work is retried rather than lost. Ch 30.

LLM-as-judge: Using a model to grade outputs against a rubric; must be validated against humans. Ch 10.

Long context: Models accepting hundreds of thousands to millions of tokens; complement, not replacement, for RAG. Ch 19.

Lost in the middle: Models use information at the start/end of the context better than the middle. Ch 19.

M (HNSW): Links kept per node in the HNSW graph. More links means better recall and more memory (about M x 2 x 4 bytes per vector) and slower construction. Ch 22.

Matryoshka (MRL): Embeddings whose leading dimensions form a good smaller embedding; enables truncation. Ch 18, 19.

MaxSim: Late-interaction score: sum over query tokens of the max similarity to any document token. Ch 19.

MCP (Model Context Protocol): Standard for exposing tools, resources and prompts to model-driven clients. Ch 19, 20.

Metadata / payload: Structured fields stored with a chunk (source, section, date, tenant) for filtering and citations. Ch 4.

Middleware (LangChain agents): Hooks around model and tool calls: limits, retries, PII, summarisation. Ch 16.

MinHash: A sketch that estimates Jaccard similarity between documents in constant space; near-duplicate detection before you pay to embed. Ch 30.

MMR (maximal marginal relevance): Retrieval that trades relevance against diversity to avoid near-duplicate chunks. Ch 9.

MRR: Mean of 1/rank of the first relevant result. App B.

Multi-query: Generating several rephrasings of the question, retrieving for each, and merging. Ch 9.

Multi-tenancy: Serving many customers from one system with strict data isolation via payload filters or shards. Ch 18.

Multivector: A Qdrant point holding many vectors (e.g. per token/patch) scored by MaxSim. Ch 19.

nDCG: Normalised discounted cumulative gain; ranking quality with position discount. App B.

Node: A LangGraph function that reads state and returns an update. Ch 14.

nprobe: The number of IVF clusters searched at query time; IVF's equivalent of ef_search, trading recall against latency. Ch 22.

Observability: Traces, logs and metrics that let you replay and diagnose a run. Ch 13.

Overlap: see Chunk overlap.

Paired permutation test: Significance test that shuffles which system produced which per-question score; correct for comparing two configs on the same eval set. Ch 21.

Parent-document / small-to-big retrieval: Retrieve small chunks for precision, return their parent section for context. Ch 9.

Payload index: Qdrant index on a metadata field to make filters fast (server mode). Ch 4, 18.

PII: Personally identifiable information; redact at ingest and on output. Ch 16, 17.

Point: A Qdrant record: id + vector(s) + payload. Ch 4.

Precision@k: Fraction of the top k that are relevant. App B.

Product quantization (PQ): Splitting a vector into subvectors and replacing each with a codebook id, compressing 10-60x at some recall cost; often combined with a rescoring pass. Ch 22.

Prompt caching: Provider discount for repeated prompt prefixes; put stable content first. Ch 17, 19.

Prompt injection: Input crafted to make the model ignore its instructions; see Indirect prompt injection. Ch 19.

Quantisation: Storing vectors in fewer bits (int8, binary, product) to cut memory; rescore with originals. Ch 18.

Query rewriting: Transforming the user question before retrieval (rephrase, expand, decompose, step back). Ch 9.

RAG: Retrieval-Augmented Generation: fetch relevant text, put it in the prompt, generate a grounded answer. Ch 1.

RAPTOR: Tree of recursive cluster summaries retrieved at all levels. App A.

Reading order: The sequence a human would read a page in. Not stored in a PDF; reconstructed from block coordinates, and wrong reading order silently corrupts every downstream chunk. Ch 28.

Recall@k: Fraction of relevant chunks in the top k. App B.

Recursion limit: LangGraph cap on steps per run; stops runaway loops. Ch 16.

Recursive character splitter: Splits on a priority list of separators (headings, paragraphs, sentences) to a target size. Ch 5.

Refusal: The system saying it cannot answer from the available information. Ch 12.

Reranker: Second-stage model reordering the top candidates by relevance. Ch 9.

Rescoring: Recomputing similarity with full-precision vectors for the top quantised candidates. Ch 18.

Retriever: A component that returns documents for a query (vector_store.as_retriever() or a function). Ch 6.

RRF (Reciprocal Rank Fusion): Fusion by Σ 1/(60 + rank). Ch 8.

Self-RAG: Model emits reflection tokens deciding retrieval and grading relevance/support. Ch 15.

Semantic cache: Cache keyed by embedding similarity of the question; returns stored answers for near-duplicates. Ch 19.

Semantic chunking: Splitting where embedding similarity between adjacent sentences drops. Ch 5.

Shard: A horizontal partition of a collection across nodes; custom shard keys for tenant isolation. Ch 18.

SimHash: A locality-sensitive fingerprint where similar documents get similar hashes, so near-duplicates can be found by Hamming distance. Ch 30.

Sparse vector: High-dimensional vector with few non-zeros (term → weight), e.g. BM25/SPLADE; stored natively in Qdrant. Ch 8.

State (LangGraph): The typed dictionary passed between nodes; reducers (e.g. add_messages) define how updates merge. Ch 14.

Step-back prompting: Asking a more general question first to retrieve background, then the specific one. Ch 9.

Structured output: Forcing the model to return a schema (Pydantic) via with_structured_output. Ch 11, 15.

Synthetic data: LLM-generated questions/answers from your corpus, used to bootstrap evals. Ch 7.

System prompt: Instructions that frame the model's role and rules; stable prefix for caching. Ch 3.

Token: Sub-word unit of text; ~4 characters of English on average. Ch 1.

Token bucket: A rate limiter that refills tokens at a fixed rate; how you respect an embedding API's tokens-per-minute limit under concurrency. Ch 30.

Tool: A function the model can call, described by name, docstring and schema (@tool). Ch 14.

ToolNode: LangGraph prebuilt node that executes the tool calls in the last AI message. Ch 15.

Trace: Recorded tree of calls (model, retriever, tools) for one run, with inputs, outputs, latency and tokens. Ch 13.

Usage metadata: Token counts returned with a model response; basis for cost accounting. Ch 13.

Vector database: Store optimised for ANN search over embeddings with filters and payloads (Qdrant). Ch 4.

Verbalised row: A table row rewritten as a sentence ('The Atlas A2 carries up to 250 kg') so that natural-language queries match it. Ch 28.

Weighted RRF: Reciprocal rank fusion with a per-retriever weight, for when you know one retriever is more reliable on your corpus. Ch 23.

Zero-shot: Prompting with instructions only, no examples. Ch 20.