Glossary

These definitions describe how the project uses each term. They are intentionally narrower than every use in the broader RAG literature.

Retrieval and generation

Ablation

A controlled evaluation that disables or changes one retrieval component so its measured contribution can be compared with the baseline.

Answer grounding

The degree to which claims in a generated answer are supported by the exact source passages supplied to the model.

Candidate pool

The bounded items returned by one retrieval layer before fusion, or the wider fused set presented to a reranker.

Chunk

The smallest stored and ranked text unit. A chunk retains its document identity, order, section path, table flag, embedding, and version stamp.

Deep profile

The retrieval profile intended for offline or agent use. It enables the model-dependent layers and uses a longer per-stage timeout.

Embedding

A fixed-length numeric vector representing text for similarity search. The default Google embedding is truncated and normalized to 1536 dimensions.

Fusion

Combining ranked lists from several retrieval layers. Sci RAG Kit uses weighted reciprocal rank fusion once per request.

HyDE

Hypothetical Document Embeddings. A model writes the kind of passage that would answer the query; the system embeds that passage and searches near it.

Interactive profile

The low-latency profile centered on vector and keyword retrieval, with shorter stage timeouts.

RAG

Retrieval-augmented generation. Evidence is retrieved from a corpus and supplied to a generator so the answer can be grounded in those sources.

Reciprocal rank fusion (RRF)

A method that scores an item from its positions in multiple ranked lists rather than comparing incompatible raw layer scores.

Reranker

A post-fusion component that reads the query and candidate text, then reorders a wider pool. It is off until an ablation justifies its cost.

Retrieval layer

One independent way to propose evidence. The five layers are vector, keyword, graph, community, and HyDE.

Retrieval scope

The rights and metadata constraints a caller applies before ranking, including license, source, year, author, journal, and DOI exclusions.

Graph and corpus

Community

A deterministic cluster of related knowledge-graph entities with a precomputed model summary and embedding.

Corpus fingerprint

Counts and embedding versions recorded on evaluation output so changes to the indexed knowledge base remain visible.

Corpus campaign

A resumable workflow that discovers DOI candidates, resolves explicit open-access rights, downloads verified direct PDFs, and writes an ingestible manifest without inferring a license from availability alone.

Corpus snapshot

A named, immutable record of per-document content hashes plus one digest, model versions, counts, and Git commit.

Entity

A canonical named concept extracted under one of the types declared in domain/domain.yaml, with retained source aliases and pointers to the chunks and documents that support it.

Knowledge graph

Typed entities and directed relationships stored as ordinary Postgres rows. It is a graph of domain concepts, not a second database.

Known retraction

A document for which an applied Crossref enrichment response contains an explicit retraction assertion. Missing enrichment is not treated as proof that a work is current.

Ontology

The allowed entity and relationship types for one scientific domain, expressed in domain/domain.yaml.

Provenance

The identity and context needed to trace evidence back to its source, including document metadata, chunk section path, and graph evidence pointers.

Relationship confidence

A calibrated knowledge-graph edge score: 1.0 for a direct statement, 0.7 for a strong implication, and 0.4 for an inference across sentences. Repeated extraction retains the highest observed score for the same typed edge.

Source bucket

An operator-defined manifest label such as journal_papers or agency_reports, usable as a retrieval allowlist.

Evaluation and operations

Blind judge pass

Grounding evaluation in which the judge sees the question, generated answer, and retrieved evidence, but not the expert reference answer.

Calibration

Comparison of model-judge labels with human labels. The project reports Cohen's kappa and agreement matrices per dimension.

Confidence interval

A bootstrap interval around a reported mean, resampling at the question level to show uncertainty from the evaluation set.

Degraded stage

A retrieval stage that timed out or errored while the rest of the request continued. Empty, skipped, and disabled stages have different meanings.

Evaluation seed question

A version-controlled question with reference documents or evidence phrases, and optionally an expert answer, used to measure the target corpus.

License class

One of the five operational redistribution categories attached to every document: public, open_commercial, open_noncommercial, restricted, or unknown.

Stage trace

Content-free operational metadata for one retrieval stage: status, duration, and candidate count. It is safe to log because it contains no query or chunk text.

Version stamp

The provider/model/dimension identity stored with an embedding so stale rows can be found and re-indexed after a model change.