Skip to content

📖 Glossary: the working vocabulary

Status: reference.

Written for you, the person writing a draft with this pipeline. Every term below appears throughout the documentation. Each entry says what the thing is to you and names the one document that owns the detail. Terms are grouped by when you first meet them, not alphabetically; read top to bottom once and the rest of the documentation gets easier.

📚 Your library, inside the pipeline

  • Citekey: the identifier your reference manager assigns each entry (sample_dt_sync_2023), which a draft cites as [@sample_dt_sync_2023]. The one hard rule of this project: a citekey may be used only if it came from your own .bib export and a real PDF was really parsed (SOUL.md).
  • Corpus: your papers as the pipeline sees them, meaning the .bib export plus the attached PDFs, synced into the ledger.
  • Ledger: the small database (content/ledger.sqlite) recording every entry: its citekey, title, whether its PDF parsed, and where the extracted text lives. Built and updated by chitragupta corpus sync; inspected with chitragupta corpus ledger (ZOTERO.md gets you here).
  • Sync: the deterministic run that reads your .bib, updates the ledger, and extracts each PDF's text. No AI is involved, so the same bibliography always yields the same citekeys.
  • Parse: turning one PDF into plain text the pipeline can search and quote. Which of the two parser backends to use is PDF-PARSER.md's question.

✍ Writing

  • Genre skill: one of the six drafting skills (survey, thesis chapter, textbook chapter, tutorial, deep research, verbatim digest). The other five work on drafts those produce: three revisers, the book assembler, and figure-drawer, which draws a figure for whichever skill hands it one. You never invoke one by name; you ask in ordinary words and the right one answers (GENRE.md).
  • Draft: the document being written, under content/drafts/.
  • Dossier: the folder written alongside a draft holding the judgment that produced it: the scope, who the reader is, which sources were kept and which were turned down, and a revision log. It is why a draft can be changed months later without re-running anything (DOSSIER.md; the walkthrough is WRITING-PROCESS.md).
  • The gate: chitragupta draft gate, the single check standing between a draft and a rendered document. It asks whether every citekey in the draft is really in the ledger. It is the only blocking check in the pipeline, and a failure sends the draft back to be rewritten, so you normally never see one.
  • Render: turning a finished draft into PDF/LaTeX/Word with its bibliography resolved. chitragupta draft render runs the gate again on the text it is about to render.
  • Stamp (the draft fingerprint): a digest of the draft's text that chitragupta draft dossier stamp records in the dossier's scope.md once the gate passes. dossier status compares against it to tell you whether the draft changed since, for example after a hand edit (USER-EDITS.md).
  • Evidence sidecar: a separate document rendered beside a draft, listing each cited source and the exact spans its dossier marked quotable. Four of the six genres produce one (DOSSIER.md).
  • Verbatim digest: the sixth genre. Private study notes written mostly in the papers' own words, each copied run closed by its citation. Never a deliverable and never a source for another draft (VERBATIM-DIGEST.md).
  • Spec, unit, registry: the book track. The spec is the book's outline, which you sign off; a unit is one chapter, generated from its own contract and accepted on its own; the registries collect terminology, claims and cross-references across accepted units (WRITE-A-BOOK.md).

🔭 Exploring

  • Topic: a group of papers about the same thing. Two kinds exist side by side: seed topics are phrases you wrote (content/seed_topics.toml), matched against your papers; emergent topics are groupings the corpus itself turned out to have, discovered by clustering (TOPIC-MODELLING.md).
  • Topic graph: how the topics relate to each other, by sharing papers and by being about similar things. chitragupta corpus discover is how you walk it (TOPIC-DISCOVERY.md).
  • Overview: the Markdown file corpus discover --out writes for one topic. It lists the topic's papers with references, its neighbours, and real sentences quoted verbatim from the papers. It is raw material for a new draft and is never a generated summary.

🔍 Checking

  • Review aid: one of the eleven read-only reports you can run over a finished draft (provenance, verbatim overlap, coverage, and so on). Every aid is advisory: it reports for you to judge and never blocks anything (REVIEW.md).
  • Provenance: the aid answering "does the cited paper actually say this?", worst matches first (CITATION-PROVENANCE.md).
  • Verbatim scan: the aid answering "how much of this wording came from my sources?" (PLAGIARISM.md).
  • Agenda: chitragupta review agenda, which merges the other aids' reports into one ranked, deduplicated worklist (AGENDA.md).
  • Unattended and surfaced: the two kinds of agenda item. An unattended item is one agenda-reviser may repair without asking (a short verbatim run, a prose finding, a citation to a paper the corpus no longer has). A surfaced item waits for you to decide.

⚙ How the machinery talks about itself

Five structural words the other documents lean on. You can write a whole book without needing them; they matter when you read the deeper pages.

  • Layer: one of the four parts of the pipeline, each with its own command: corpus (deterministic), drafting (generative, reviewed by you), enrichment (optional depth), review (advisory reports). The remaining commands (init, install, doctor, figure) set up a project or work on one figure, and sit outside the layers. FEATURES.md draws them.
  • Stage: one step of the enrichment layer (docling, embed, bertopic, extract-keywords, seed-topics, converge, topic-graph). Each reports its own honest status: ok, partial, skipped or error. LADDERS.md states which of the four changes the run's exit code.
  • Artefact: a file a stage or a sync writes under content/, which later commands read. The layers never call each other. They leave files for each other, which is why you can run any piece alone.
  • Ladder and tier: the two ways the pipeline handles having more than one way to do a job. A ladder falls back automatically and tells you which rung it landed on; a tier is a menu you chose from, and stops rather than substituting when your choice is unavailable (LADDERS.md).
  • Enrichment / the enrich extra: the optional heavy machinery (layout-aware parsing, semantic search, topic modelling), installed with pip install 'chitragupta-cli[enrich]'. Nothing else needs it. Everything that can use it says so when it is absent instead of failing.