citegeist

citegeist is a research-oriented bibliography workbench for building, expanding, and auditing BibTeX libraries.

The aim is not just to store citations. The aim is to help with the harder problem: finding, improving, connecting, and checking the literature around a topic while keeping BibTeX as a first-class output format.

Repo Description

citegeist is a BibTeX-native research tool for citation extraction, metadata enrichment, citation-graph expansion, and semantic search over scholarly sources.

Scope

The project is intended to support a workflow like this:

1. Start from rough references extracted from papers, notes, syllabi, or dissertations.
2. Convert them into draft BibTeX entries.
3. Enrich and correct those entries using external scholarly metadata sources.
4. Persist entries, identifiers, abstracts, and citation edges in a local database.
5. Traverse the citation graph outward to discover additional relevant works.
6. Search the local corpus semantically using abstracts and extracted full text.
7. Export verified results back into BibTeX for LaTeX use.

Why A New Codebase

This repository starts cleanly rather than extending the older bib/ toolkit directly.

The older toolkit is useful as prior art:

• it demonstrates identifier-driven metadata augmentation;
• it caches PDFs and extracted plaintext;
• it shows one workable model for bibliography growth.

But it is not the right long-term base:

• it is Python 2-era code;
• it is shell-script centric;
• it does not provide a normalized database for graph workflows;
• it is not structured as a reusable Python 3 library.

citegeist keeps the useful ideas and rebuilds the foundation around a cleaner Python 3 package boundary.

Current Status

The initial repo includes:

• pybtex-backed BibTeX parsing and export in a repo-local virtual environment;
• a SQLite-backed bibliography store;
• a small CLI for ingest, search, inspection, and export;
• review-state tracking on entries, per-field ingest provenance, and field-level conflict review;
• plaintext reference extraction into draft BibTeX for numbered, APA-like, wrapped-line, and simple book-style references;
• staged plaintext reference extraction that now preserves more structured metadata from legacy references, including year suffixes, identifiers, volume/issue/pages, and thesis/report/web-style venue hints;
• a reference-extraction backend seam with the local heuristic parser as the default implementation, so optional external backends can be added later without changing the core extract workflow;
• standalone verification and disambiguation of free-text references or partial BibTeX into auditable BibTeX/JSON results with x_status, x_confidence, x_source, x_query, and alternate-candidate traces;
• identifier-first metadata resolution for DOI, PMID/PubMed, OpenAlex, DBLP, arXiv, and DataCite-backed entries, with OpenAlex/DataCite/PubMed title-search fallback;
• local citation-graph traversal over stored cites, cited_by, and crossref edges;
• Crossref- and OpenAlex-backed graph expansion that materializes draft related works and edge provenance;
• a dedicated source-client layer with fixture/cache support for live-source development;
• OAI-PMH Dublin Core harvesting for institutional repositories and thesis/dissertation sources;
• OAI-PMH repository discovery via Identify, ListSets, and ListMetadataFormats to target harvests more precisely;
• bibliography bootstrap workflows that can start from a seed .bib, a topic phrase, or both;
• batch bootstrap orchestration from JSON job files containing seed BibTeX paths, topic phrases, or both;
• normalized tables for entries, creators, identifiers, and citation relations;
• full-text-search-ready indexing over title, abstract, and fulltext when SQLite FTS5 is available;
• tests covering parsing, ingestion, relation storage, and search.

Example applications live alongside the core package rather than defining it. Current examples include:

• a comprehensive CLI cookbook in examples/cli/README.md;
• a topic-only bootstrap workflow for artificial life in examples/artificial-life/README.md;
• a browser-oriented literature explorer demo with a small HTTP bridge, static HTML/JS shell, and lightweight graph view in examples/literature-explorer/README.md;
• the TalkOrigins bibliography pipeline under citegeist.examples.talkorigins with a usage guide in examples/talkorigins/README.md.

The prioritized execution plan lives in ROADMAP.md.

Status Assessment

citegeist is no longer just a storage-and-export skeleton. It now covers the main early pipeline the project set out to make usable on one local machine:

1. ingest or extract rough references,
2. verify and improve them before trust is assumed,
3. normalize and store them with provenance,
4. expand outward through citation links,
5. review and export BibTeX again.

In practical terms, the strongest implemented areas are now:

• BibTeX-native local storage, review state, provenance tracking, and export;
• rough-reference handling through both heuristic extraction and standalone verification/disambiguation;
• conservative metadata enrichment and citation-graph expansion from scholarly APIs;
• fixture-backed parser and source-client workflows that keep improvement work auditable;
• a lightweight local demo surface for topic discovery, topic expansion, extraction, verification, and graph inspection in the browser without introducing a full web framework dependency.

The main gaps are no longer the basic pipeline. The main gaps are evaluation depth and researcher ergonomics: broader regression fixtures, clearer comparative quality measurement across parsers/resolvers, stronger review workflows for larger corpora, and a richer review UI than the current demonstration shell.

Positioning

Compared with other bibliographic tooling, citegeist is strongest when bibliography work starts messy and needs to become structured:

• Against reference managers such as Zotero or JabRef, citegeist is currently weaker as a polished day-to-day library manager or sync-oriented desktop app, but stronger as a BibTeX-first command-line workbench for extraction, repair, provenance, and graph-oriented discovery.
• Against parser-focused tools such as AnyStyle, GROBID, or ParsCit-style systems, citegeist is not trying to outcompete them as a dedicated citation parser. Instead, it now uses their staged-parsing ideas and can optionally call external parsers while keeping a local default parser and a normalized BibTeX-oriented downstream workflow.
• Against verifier/disambiguator workflows like VeriBib, citegeist now covers the same high-value pre-ingest verification pattern, but places it inside a broader local pipeline that also stores results, resolves identifiers, expands citation graphs, and exports reviewed BibTeX.
• Against process-heavy corpus updaters like TOA-Bib-Updater, citegeist now adopts the useful operational pattern of staged artifacts and reviewable outputs, but keeps the core centered on reusable library/database primitives rather than one source-specific acquisition script.

The clearest differentiator at this point is integration. citegeist is becoming a local bibliography workbench that combines extraction, verification, enrichment, graph expansion, topic-aware review, and BibTeX export in one toolchain rather than treating those as unrelated utilities.

Layout

citegeist/
  src/citegeist/
    bibtex.py
    examples/
    storage.py
  tests/
    test_storage.py
  pyproject.toml

Quick Start

cd citegeist
python3 -m virtualenv --always-copy .venv
.venv/bin/pip install -e .
.venv/bin/pip install pytest
mkdir -p .cache/citegeist
PYTHONPATH=src .venv/bin/python - <<'PY'
from citegeist import BibliographyStore

bib = """
@article{smith2024graphs,
  author = {Smith, Jane and Doe, Alex},
  title = {Graph-first bibliography augmentation},
  year = {2024},
  abstract = {We study citation graphs for literature discovery.},
  references = {miller2023search}
}

@inproceedings{miller2023search,
  author = {Miller, Sam},
  title = {Semantic search for research corpora},
  year = {2023},
  abstract = {Dense retrieval improves recall for academic search.}
}
"""

store = BibliographyStore("library.sqlite3")
store.ingest_bibtex(bib)
print(store.get_relations("smith2024graphs"))
print(store.search_text("semantic"))
store.close()
PY
.venv/bin/python -m pytest -q

Or use the CLI directly:

cd citegeist
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 ingest references.bib
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 search "semantic search"
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 search "origin" --topic abiogenesis
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 show --provenance --conflicts smith2024graphs
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 set-status smith2024graphs reviewed
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 resolve-conflicts smith2024graphs title accepted
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 apply-conflict smith2024graphs title
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 bootstrap --seed-bib seed.bib --topic "bayesian nonparametrics"
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 bootstrap --topic "bayesian nonparametrics" --preview --topic-commit-limit 5
PYTHONPATH=src .venv/bin/python -m citegeist extract references.txt --output draft.bib
PYTHONPATH=src .venv/bin/python -m citegeist extract references.txt --backend heuristic --output draft.bib
PYTHONPATH=src .venv/bin/python -m citegeist compare-extract references.txt --backend heuristic --backend anystyle --output compare.json
PYTHONPATH=src .venv/bin/python -m citegeist compare-extract references.txt --backend heuristic --backend grobid --summary --output compare-summary.json
PYTHONPATH=src .venv/bin/python -m citegeist compare-extract references.txt --backend heuristic --backend grobid --summary --max-rows-with-differences 0 --output compare-check.json
PYTHONPATH=src .venv/bin/python -m citegeist verify --string '"Graph-first bibliography augmentation" Smith 2024' --context "citation graphs" --format json
PYTHONPATH=src .venv/bin/python -m citegeist verify --string 'Evans 1960' --context "bottlenose dolphin echolocation" --llm --llm-base-url http://localhost:11434 --llm-model qwen3 --llm-role both --format json
PYTHONPATH=src .venv/bin/python -m citegeist verify --bib draft.bib --output verified.bib
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 resolve smith2024graphs
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 resolve-stubs --doi-only --preview --limit 25
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 resolve-stubs --doi-only --all-misc --limit 25
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 sync-jabref my-library.bib --output my-library.enriched.bib
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 topics
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 topic-entries abiogenesis
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 export-topic abiogenesis --output abiogenesis.bib
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 graph smith2024graphs --relation cites --depth 2 --missing-only
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 graph smith2024graphs --relation cites --depth 2 --format json-graph
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 graph smith2024graphs --relation cites --depth 2 --format dot
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 graph smith2024graphs --relation cites --depth 2 --format dot --output graph.dot
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 graph smith2024graphs --relation cites --depth 2 --format json-graph --output graph.json
PYTHONPATH=src .venv/bin/python -m citegeist graph-view graph.json --output graph.html --title "CiteGeist Graph"
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 expand smith2024graphs --source crossref
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 expand smith2024graphs --source openalex --relation cited_by --limit 10
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 expand-topic abiogenesis --topic-phrase "abiogenesis origin chemistry" --source openalex --relation cites --seed-key seed2024 --min-relevance 0.3 --preview
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 set-topic-phrase abiogenesis "abiogenesis origin chemistry prebiotic"
PYTHONPATH=src .venv/bin/python -m citegeist discover-oai https://example.edu/oai
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 harvest-oai https://example.edu/oai --metadata-prefix mods --from 2024-01-01 --until 2024-12-31 --limit 10
PYTHONPATH=src .venv/bin/python -m citegeist --db library.sqlite3 export --output reviewed.bib

For a fuller option-by-option CLI cookbook, see examples/cli/README.md.

Broad BibTeX exports skip DOI-only placeholder records such as Referenced work N by default. Use --include-stubs on export or export-topic if you want those entries included anyway.

Long-running CLI commands report progress on stderr so stdout remains clean for JSON, BibTeX, or tabular output.

BibTeX parse/render round-trips normalize simple escaped special characters such as \_, \&, and \% back to plain field values internally, then re-escape them on export. This prevents repeated commands such as resolve from turning a valid field like discovered\_from = {...} into discovered\\_from = {...} after rewriting an entry.

Crossref reference expansion is intentionally conservative about weak discoveries. If a cited reference has no DOI and Crossref only exposes it as an unstructured citation blob, expand --source crossref, expand-topic --source crossref, and bootstrap flows now skip materializing that record unless the fallback metadata looks like a cleaner non-misc work such as conference proceedings. When Crossref does expose thesis or dissertation references only as unstructured text, citegeist now also tries to extract the actual work title instead of keeping the entire ProQuest-style citation blob in the title field. This reduces junk @misc entries and cleaner @phdthesis fallbacks whose title field is really a pasted citation string.

OpenAlex expansion is also conservative about noisy secondary records. Discoveries now prefer DOI-based citation keys when a DOI is present, which reduces parallel doi... and openalex... entries for the same work. OpenAlex abstract imports are sanitized to drop obvious webpage/export blobs, and DOI-less records are filtered when they look like venue-title stubs, generic container records, or weak review-like shadows of an already present book, chapter, proceedings paper, or dissertation with the same title. Both OpenAlex and Crossref discovery paths also normalize some malformed author strings from upstream metadata, including inverted-initial patterns such as J., Fogel L., into stable BibTeX names like Fogel, L. J.. Preview mode uses the same admission rules as write-enabled expansion, so rejected candidates disappear before you commit them to the store.

For live-source development, prefer fixture-backed or cache-backed source clients so resolver and expansion work can be exercised repeatedly without re-hitting upstream APIs on every run.

JabRef Workflow

citegeist is not a replacement for JabRef's day-to-day BibTeX editing UX. The intended near-term integration model is file-based round-tripping:

• use JabRef to inspect, edit, and review your main .bib library;
• use citegeist sync-jabref to ingest that file into CiteGeist, run metadata enrichment against imported entries, and write an enriched .bib export back out;
• reopen the enriched file in JabRef for human review and curation.

That keeps JabRef as the primary manual review surface while letting CiteGeist handle source-backed resolution and discovery work that reference managers usually do not automate well.

Useful options for that round trip:

• --in-place: overwrite the input .bib file instead of writing to a separate export path
• --annotate-review: add x_citegeist_* sidecar fields such as review status, open-conflict count, and last source label so JabRef can surface CiteGeist review cues directly in the BibTeX record
• --no-resolve: skip live metadata resolution and only perform import plus re-export

Adopted Ideas From Earlier Repos

citegeist now absorbs two useful patterns from adjacent bibliography tools while keeping them inside the main Python 3 package boundary:

• From VeriBib: a standalone verify workflow for ambiguous strings or rough BibTeX, with explicit confidence/status audit fields and alternate-candidate traces before you commit changes to the main library.
• From TOA-Bib-Updater: resumable, artifact-oriented corpus processing remains the preferred process model for large imports. In practice this already appears in the TalkOrigins example pipeline through saved manifests, review exports, duplicate reports, and staged topic-phrase review flows.

Those ideas are now implemented enough to matter operationally, not just directionally: VeriBib's main contribution has become a core verify command, while TOA-Bib-Updater's main contribution remains process shape and review artifacts rather than parser or storage internals.

Parsing Sources

The plaintext-reference parser is still local and heuristic-first, but its current direction explicitly borrows ideas from earlier citation-parsing work:

• Conceptual influence: GROBID's staged parsing model and bibliographical-reference annotation guidance informed the split between reference-block segmentation, field extraction, and metadata recovery, especially for identifiers, year variants, and venue/page structure.
• Conceptual influence: AnyStyle and ParsCit informed the emphasis on treating reference parsing as a separable stage with gold-fixture-driven improvement rather than a one-pass punctuation split.
• In-repo code prior art: some of the newer heuristics were adapted from existing CiteGeist code in src/citegeist/talkorigins.py and src/citegeist/expand.py, particularly around entry-type guessing, fragment cleanup, and thesis/report handling for citation-like blobs.

External references used for the current parser direction:

• GROBID principles and parsing architecture: https://grobid.readthedocs.io/en/latest/Principles/
• GROBID bibliographical-reference annotation notes: https://grobid.readthedocs.io/en/latest/training/Bibliographical-references/
• AnyStyle project: https://github.com/inukshuk/anystyle
• ParsCit paper: https://aclanthology.org/L08-1291/

The built-in extraction backends are:

• heuristic: the default local parser, always available
• anystyle: an optional adapter around the AnyStyle CLI when anystyle is installed locally
• grobid: an optional adapter around a running GROBID service using /api/processCitationList

The backend interface exists so future GROBID- or other parser adapters can be registered without replacing the local parser or changing the CLI contract.

LLM-Assisted Verify

citegeist verify can optionally use a local LLM for two bounded tasks:

• expand: infer missing bibliographic clues from free text and context
• rerank: advisory reranking of already fetched resolver candidates

Example:

PYTHONPATH=src .venv/bin/python -m citegeist verify \
  --string 'Evans 1960' \
  --context "bottlenose dolphin echolocation" \
  --llm \
  --llm-base-url http://localhost:11434 \
  --llm-model qwen3 \
  --llm-role both \
  --format json

Supported local endpoint styles:

• OpenAI-compatible APIs such as http://localhost:11434/v1
• Ollama native chat APIs such as http://localhost:11434

For the current local GenieHive setup, this also works directly:

PYTHONPATH=src .venv/bin/python -m citegeist verify \
  --string 'Evans 1960' \
  --context "bottlenose dolphin echolocation" \
  --llm \
  --llm-base-url http://127.0.0.1:8800/v1 \
  --llm-api-key change-me-client-key \
  --llm-model general_assistant \
  --llm-role both \
  --format json

There is also a local smoke script for the LLM helper path alone:

make live-verify-llm-smoke

Safety constraints:

• the LLM is never trusted for DOI or identifier invention
• the LLM only fills missing query clues or suggests candidate order
• exact status still requires verified resolver evidence, not LLM output
• if the LLM fails or returns unusable JSON, verify falls back to the normal resolver-only path

To compare backend output on the same plaintext references, use compare-extract. It aligns entries by ordinal/reference block and emits JSON with per-backend payloads plus a differing_fields summary for each row. Add --summary when you want a compact evaluation artifact with disagreement counts by field and backend presence counts instead of the full row-by-row payload. Add --max-rows-with-differences and/or --max-field-difference-count when you want CI-style failure thresholds; the command will emit the summary JSON and return a nonzero exit code if the limits are exceeded.

For regression-oriented parser work, keep a small curated plaintext fixture set and run compare-extract against multiple backends before changing heuristics. That makes backend disagreement explicit and gives you a stable review artifact for parser changes.

For the optional AnyStyle backend, install the CLI separately and then run:

PYTHONPATH=src .venv/bin/python -m citegeist extract references.txt --backend anystyle --output draft.bib

If the binary is not on PATH, set CITEGEIST_ANYSTYLE_BIN=/path/to/anystyle. If you want a custom AnyStyle parser model, set CITEGEIST_ANYSTYLE_PARSER_MODEL=/path/to/model.mod.

For the optional GROBID backend, start a GROBID service and then run:

PYTHONPATH=src .venv/bin/python -m citegeist extract references.txt --backend grobid --output draft.bib

By default CiteGeist targets http://127.0.0.1:8070. Override that with CITEGEIST_GROBID_URL=http://host:port if your service is elsewhere.

Example Application

• Use stage-topic-phrases to load those suggestions into the database as review items. Staging stores the candidate in suggested_phrase and marks the topic pending without changing the active expansion_phrase.

• Use export-topic-phrase-reviews to write an editable JSON template directly from the database for the currently staged suggestions. That gives you a round-trip path from DB review queue to file edits and back into review-topic-phrases.

• Use review-topic-phrase to accept or reject one staged suggestion in place. Accepting a suggestion copies it into expansion_phrase and clears it from the staged review queue; rejecting it preserves the staged suggestion together with its review state.

• Use review-topic-phrases when you want to apply many accept/reject decisions from one JSON file. Each item should carry slug, status, and optional phrase / review_notes.

• Use apply-topic-phrases when you want a direct patch path instead of the staged review flow. It accepts either the raw suggestion list or an object with a topics list, and will apply suggested_phrase or phrase to matching topic slugs immediately.

• Use topic-phrase-reviews --phrase-review-status pending when you want a compact audit view of unresolved staged suggestions, including both the current live phrase and the pending replacement.

• Use enrich-talkorigins when you want to target those weak canonical entries for resolver-based metadata upgrades before retrying graph expansion on imported topic slices.

• Use review-talkorigins when you want one JSON review artifact that combines weak canonical clusters with dry-run enrichment outcomes for manual cleanup.

• Use expand-topic when you already have both a topic phrase and a curated topic seed set in the database: it expands outward from the topic’s existing entries, then only assigns discovered works back to that topic if they clear a topic-relevance threshold. Write-enabled assignment is stricter than preview ranking: a candidate must clear the score threshold and show a non-generic title anchor to the topic phrase, so broad methods papers do not get attached just because their abstracts or related terms overlap. On large noisy topics, prefer --seed-key to restrict the run to just the trusted seed entries you want to expand from, and use --preview first to inspect discovered candidates and relevance scores before writing anything.

• Use set-topic-phrase to store a curated expansion phrase on the topic itself. When a stored phrase exists, expand-topic will use it automatically if you do not pass --topic-phrase. Batch bootstrap jobs can also set topic_slug, topic_name, and topic_phrase so curated topic metadata is created as part of the run.

• Use topics --phrase-review-status pending when you want to audit only topics whose staged phrase suggestions still need review.

• --allow-unsafe-search-matches exists only for bounded experiments on copied databases when you explicitly want to relax trust to exercise downstream expansion behavior.

The TalkOrigins corpus pipeline remains in the repository as an example application rather than a core package surface. Use the example-scoped Python namespace:

from citegeist.examples.talkorigins import TalkOriginsScraper

and the example-scoped CLI commands:

PYTHONPATH=src .venv/bin/python -m citegeist example-talkorigins-scrape talkorigins-out --limit-topics 5 --limit-entries-per-topic 20
PYTHONPATH=src .venv/bin/python -m citegeist example-talkorigins-validate talkorigins-out/talkorigins_manifest.json
PYTHONPATH=src .venv/bin/python -m citegeist example-talkorigins-duplicates talkorigins-out/talkorigins_manifest.json --limit 20 --preview --weak-only

The older scrape-talkorigins-style command names remain available as compatibility aliases. The full example workflow and reconstruction notes live in examples/talkorigins/README.md.

For a smaller example that starts from a topic phrase alone, see examples/artificial-life/README.md.

Correction files are simple JSON:

{
  "corrections": [
    {
      "key": "smith jane|1999|weak duplicate",
      "entry_type": "article",
      "review_status": "reviewed",
      "fields": {
        "journal": "Journal of Better Metadata",
        "doi": "10.1000/weak",
        "note": null
      }
    }
  ]
}

fields values overwrite the canonical entry for that duplicate-cluster key. Set a field to null to remove it.

Live-source workflow:

cd citegeist
export CITEGEIST_SOURCE_CACHE=.cache/citegeist
export CITEGEIST_LIVE_TESTS=1
PYTHONPATH=src .venv/bin/python -m pytest -m live -q
PYTHONPATH=src .venv/bin/python scripts/live_smoke.py

By default, live tests are skipped. They only run when CITEGEIST_LIVE_TESTS=1 is set.

Convenience targets:

make test
make test-live
make live-smoke

Near-Term Priorities

• source adapters beyond OAI-PMH for additional non-DOI scholarly ecosystems.

See ROADMAP.md for the prioritized phase plan and rationale.

Naming

The name is intended to be short, distinct, and memorable:

• cite for citation work;
• geist for the organizing intelligence around the literature.