Islam West Africa Collection (IWAC) MCP Server

Get one audiovisual record by id: full description and transcription (where one exists), creator/publishing channel, duration, medium, subjects, places, language, rights, source, and three distinct links — `url` (the IWAC page, the one to cite), `external_url` (where a harvested video plays) and `media_url` (a deposited file). `source_type` says which to expect.

List audiovisual materials, newest first (francophone web video from Burkina Faso, Togo and Benin; deposited Nigerian Hausa/Arabic recordings). Filter by country, publishing channel or `source_type`.

Search audiovisual materials by keyword and metadata: francophone web video from Burkina Faso, Togo and Benin (TV reports, association and campus recordings), plus deposited Nigerian Hausa/Arabic recordings. Keyword matches title, creator, publisher, subject, spatial, language, source, the item's own description (the richest text most of these items have) and its transcription where one exists. Each row says which population it is from (`source_type`) and carries either `external_url` (a video to watch) or `media_url` (a file), never both.

Get full details of an index entry by id (raw dataset columns, French names — Titre, Prénom, Coordonnées…).

Search the IWAC authority index (persons, places, organisations, events, subjects) by name. Accent/case-insensitive.

Aggregate AI polarity, centrality and subjectivity across a filter set. 5 models scored the corpus independently — gpt-5-6-luna, mistral-small-2603, deepseek-v4-flash-0731, gemma-4-31b-it, qwen3-8-27b — so model:"all" returns each one's distribution plus how often they AGREE. Treat disagreement as a fact about the judgement rather than noise: corpus-wide the panel is unanimous on polarity for only ~32% of articles, so in a set where the models split no single one's number should be quoted alone. All three scales are ordinal French labels; subjectivity is much the weakest and ships a caveat to quote with it. Articles were scored whether or not their full text ships, so these shares are not subject to the OCR coverage limit. The models do NOT all cover the same articles, so read each one's `coverage` before comparing counts: ~51 non-francophone articles are unscored by design, and qwen3-8-27b is 200 further short on articles peripheral to Islam.

Filter articles by gpt-5-6-luna sentiment labels (accent/case-insensitive exact match). One model's reading, not a consensus — 4 other models scored the same articles and often disagree; get_sentiment_distribution with model:"all" shows by how much. `subjectivity` is much the weakest of the three scales, so treat a set selected on it as a lead to read rather than as a finding.

Get one article (by id): full metadata, the AI abstract (description_ai), AI sentiment, and OCR text. Pass a `keyword` to get ~2000-char excerpts around each match instead of the full (capped) OCR.

Search IWAC newspaper articles by keyword (title + OCR + AI abstracts, French and English), country, newspaper, subject, and date range. Use French concept keywords regardless of the user's report language. Matching is accent- and case-insensitive.

Overall statistics for every IWAC subset, including `fulltext_coverage` — how many items in each subset actually carry searchable full text in this public dataset. Read that before treating any keyword count as a full-text census.

How often the top values of a multi-valued field appear on the SAME item — a subject/place co-mention matrix. Answers 'what is X discussed alongside' without reading anything: the pair counts are the structure of the tagging. Returns the top values, the full symmetric matrix (diagonal = each value's own count) and the strongest pairs.

Compare article counts, newspaper counts, date ranges, and gpt-5-6-luna polarity across countries.

Get one archival document (by id): full metadata, AI description, and OCR text. Pass a `keyword` to get ~2000-char excerpts around each match instead of the full (capped) OCR — useful for long documents.

Search the small archival-documents subset (~26 items: Islamic association reports, flyers, project documents — mostly Burkina Faso). Use French concept keywords regardless of the user's report language. Most have OCR text and an AI description. Call with no arguments to list all.

Retrieve the full text and metadata of one IWAC item by an id returned from `search` (format '<category>:<number>', e.g. 'articles:28576'). Returns {id, title, text, url, metadata}: `text` is the item's OCR / abstract / transcription / description, `url` is the canonical islam.zmo.de link to cite, and `metadata` holds the remaining fields (author, date, country, newspaper, AI sentiment, …). Categories: articles, publications, references, documents, index, audiovisual, images.

Rank the values of one multi-valued field across a filtered set — the direct way to answer 'which places does this coverage name most', 'who signs these articles', 'what subjects dominate'. Pipe-joined fields (subject, spatial, author, language, country) are split, so an article tagged 'Prière|Ramadan' counts once for each. Optional over_time adds the per-year share of items that carry ANY value for the field, which is how you see e.g. bylines appearing as the press professionalises.

Get one photograph by id: title, photographer, capture date, place and coordinates, subjects, rights, the IIIF manifest, and the full-resolution `image_url`. The server returns URLs, not image bytes.

Search the IWAC photographs (30 items: mosques, radio stations, schools, signage and street scenes documented during fieldwork). Keyword matches title, creator, subject, place and the rare caption. Each result carries `image_url` (the full-resolution file), `coordinates` ('lat, lng' where known) and the canonical IWAC page. Call with no arguments to list all. Captions are almost never present, so prefer subject/place filters over keywords, or semantic_search_images when it is enabled.

Readability, lexical richness and length of the press text, averaged by year, newspaper or country. `Lisibilite_OCR` is a French readability score (higher = easier); `Richesse_Lexicale_OCR` is MATTR, a moving-average type-token ratio that is ALREADY length-robust — do not normalise it by word count or bin it by length. Readability is computed against a French lexicon, so non-French items are excluded from that metric (and counted in readability_excluded) rather than reported as unreadable; MATTR and word count need no lexicon and cover everything. Only items whose full text ships in this public dataset carry these columns at all.

List lieux from the IWAC index, sorted by frequency (most-referenced first). The optional 'country' filter selects entries that APPEAR IN records from that country (mentioned-in, not located-in), ranked by collection-wide 'frequency' — so foreign and cross-border entries can appear. Nigeria returns none here (index frequency is computed from articles + publications + references, which have no Nigerian items — Nigeria is audiovisual only).

Per-newspaper article counts and date ranges.

List the Islamic periodical/series titles in the publications subset, with issue counts and year ranges. Use the returned newspaper value as the `newspaper` filter on search_publications.

List personnes from the IWAC index, sorted by frequency (most-referenced first). The optional 'country' filter selects entries that APPEAR IN records from that country (mentioned-in, not located-in), ranked by collection-wide 'frequency' — so foreign and cross-border entries can appear. Nigeria returns none here (index frequency is computed from articles + publications + references, which have no Nigerian items — Nigeria is audiovisual only).

Places named by a filtered set of items, joined to the index's authority records so each carries coordinates where the index has them. Use this rather than get_field_distribution when the question is geographic — where coverage clusters — and the plain ranking when it is not. Only `Lieux` index entries are geocoded (555 of 683); persons, organisations and events carry no coordinates and never will, and any named place with no index entry comes back under `ungeocoded` rather than being dropped.

Full OCR text of a publication, optionally returning ~2000-char excerpts around keyword matches (accent-insensitive; capped — see match_count vs excerpts_returned).

Search Islamic publications (periodical issues, books). `keyword` matches title, subject, table of contents, and full OCR text (TOC hits come back as matching_toc_entries); use French concept keywords regardless of the user's report language. Filter by newspaper/series, subject, country and year. Use list_periodicals to discover series titles, and get_publication_fulltext for keyword excerpts from a single issue.

Full bibliographic record for one academic reference (by id), including the complete abstract (present for ~51% of references), subjects, DOI/URL, and host-work details (book, volume, issue, pages).

Search academic references (journal articles, book chapters, theses, books, reports) by keyword and metadata. `keyword` is a single substring match over title + abstract, so search ONE term per call (combined terms like 'pèlerinage Mecque' miss results). References are multilingual: try French and English title/abstract keywords when relevant; metadata/filter values such as `reference_type` and `language` use French labels. Results include a short abstract snippet — use get_reference for the full abstract and bibliographic detail.

Search the Islam West Africa Collection across newspaper articles, Islamic publications, archival documents, academic references, audiovisual recordings, photographs, and the authority index (persons/places/organisations/events/subjects). Pass ONE concept or name — e.g. 'Tijaniyya', 'laïcité', 'Sheikh Gumi', 'pèlerinage'. Matching is accent- and case-insensitive; a multi-word query requires every word to appear somewhere in the item, so prefer a single concept per call. Write query strings and concept keywords in French for press/publication/document/index discovery even when the user's report language is not French. Academic references are multilingual, so try French and English title/abstract terms when relevant; metadata/filter labels remain French. Use the French transliteration of Islamic terms (Tabaski not 'Eid al-Adha', charia not 'sharia', Maouloud not 'Mawlid'). Returns {results:[{id,title,url,category}], ranking}; each result's `category` names its subset and the `ranking` field documents the ordering. Pass an id to `fetch` to read the full text. For filtered queries (by country, date, or newspaper) use the search_* tools instead.

A 2-D scatter of a filtered set, projected from the stored 768-dimension embeddings by PCA. Shows which items sit near each other in meaning — where a set splits into distinct strands and where it is one cloud. Read `explained_variance` before drawing any conclusion: with 768 dimensions the first two components usually carry a modest share, and a scatter explaining 6% of the variance is a much weaker claim than one explaining 40%. This is PCA, not UMAP: it spreads the broadest axes of variation and flattens fine cluster structure, so it is not comparable to the semantic landscapes on islam.zmo.de. Needs no API key — the vectors are a column in the dataset — but only items whose full text ships are embedded at all. NOTE the payload scales with `limit`: a point cloud is a chart, not something a text-only client can read, so for those the useful part is the explained-variance summary rather than the coordinates. Keep `limit` low unless a chart is going to be drawn.

The items nearest to a given one in meaning, by cosine similarity over the stored embeddings. Answers 'what else is like this' without a keyword — it finds pieces on the same event or theme that share no vocabulary. A neighbour above ~0.85 is usually the same story reprinted or lightly rewritten, which is how to spot syndication in this corpus; 0.6-0.8 is 'same subject, different piece'. Needs no API key: the item's own vector is a column, so nothing has to be embedded at request time. This is per-item, NOT the corpus-wide near-duplicate sweep — that is an all-pairs job and belongs offline.

List sujets from the IWAC index, sorted by frequency (most-referenced first).

Counts of matching items per year (or month) — the direct way to chart coverage trends over time instead of paging through search results. Defaults to articles; also works on publications, references, documents, audiovisual, and images. Accepts the same filters as the corresponding search_* tool (keyword = ONE substring over the subset's text fields, country, newspaper/series, subject, date range). Optional group_by=country|newspaper returns one distribution per group. Items dated only to a year keep a bare-year key even at month granularity; undated items are counted in undated_count, never dropped silently. Set calendar=hijri to bucket by the Islamic (Umm al-Qura) calendar instead — with granularity=lunar_month this collapses every year into the twelve lunar months, which is the ONLY way to see observance-driven coverage (Ramadan, Dhu al-Hijja/hajj, Shawwal/Korité): the lunar year drifts ~11 days against the Gregorian, so a Gregorian axis smears each observance across all twelve months. Hijri buckets need a full YYYY-MM-DD, so items dated only to a year or month are reported in imprecise_date_count.

How a filtered set distributes across the precomputed LDA topics, each labelled by its top terms (articles carry 30 topics and are ~99.5% classified; references have their own 33-topic model and only ~46% carry an assignment, so read its `classified` against `total_matches`). Topics are assigned offline over the full text, so they describe what a piece is ABOUT rather than which words it contains — use this instead of keyword counting to map a corpus. Optional over_time returns per-year counts for the leading topics. min_prob keeps only articles where the topic is at least that dominant (mean assignment probability is 0.34, so 0.5 is already a strong filter).