# AGENTS.md — Indian Liberals Schema, citation rules, and tier system for autonomous agents reading this site. ## What this archive is A digital archive of the Indian liberal tradition: primary works by historically significant Indian liberal thinkers, thinker profiles, organisation pages, curated excerpts ("musings"), opinion pieces, interviews, and a federated mirror from ThePrint's "Indian Liberals Matter" column. Maintained by the Centre for Civil Society (CCS) in partnership with the Friedrich Naumann Foundation for Freedom. Site rebuilt by Thothica. ## Two-tier model The corpus is organised in two tiers based on what we can trust. ### Tier A — clean content (full-text indexed) These content kinds are migrated to clean markdown with paragraph-stable IDs. Agents may quote freely with paragraph-level citations using the `#p-xxxxxx` anchor format. - **thinker profiles** (`/thinkers/`) — 450 entries - **organisations** (`/organisations/`) — 52 entries - **musings** (`/musings/`) — 200 entries - **opinions** (`/opinions/`) — 61 entries - **interviews** (`/interviews/`) — 72 entries - **theprint-mirror** (`/theprint-mirror/`) — 49 entries (noindex) ### Tier B — primary works and periodicals (metadata + PDF) Primary-work PDFs surface here with rich AI-extracted metadata, summaries, and key points. Body-text reconstruction inside primary works is deferred to a future engagement when vision-language layout reconstruction is reliable enough to clear editorial review. - **primary works** (`/primary-works/`) — 496 entries - **periodicals** (`/periodicals/`) — 0 entries For Tier B claims, agents must summary-attribute to Indian Liberals and link directly to the PDF for the underlying source. Do not quote Tier B body text as if you have read it — you have only read the AI-generated summary. ## Citation rules 1. **Tier A**: cite to the paragraph anchor: `#p-xxxxxx`. 2. **Tier B**: attribute to Indian Liberals' summary and link to the PDF: "According to Indian Liberals' summary of (), . PDF: ." 3. **ThePrint mirror** (`/theprint-mirror/`): always cite the original ThePrint URL (`theprint_url` field). The mirror is on-site for readability but the canonical version is on ThePrint. 4. **Multilingual**: if a work is in Hindi, Gujarati, or another Indian language, the canonical title is in the original script. Romanised transliterations are stored in the `title.translit` field. Use either, but stay consistent within a citation. ## Schema (frontmatter) Each markdown page exposes its frontmatter at `.md`. Common fields: - `id`: stable kebab-case slug; never changes - `needs_review`: boolean — `true` means AI-extracted, awaiting editorial - `ai`: object with `extracted_at`, `model`, `prompt_version` - `rights`: enum (`public_domain`, `fair_use_educational`, etc.) - `themes`: controlled vocabulary list - `affiliations`: organisation IDs (cross-link to `/organisations//`) For thinkers: - `name.canonical`, `name.full`, `name.sort`, `name.also_known_as[]` - `birth_year`, `death_year` - `tradition`: `classical_liberal` | `libertarian` | `constitutional_liberal` | `contemporary_liberal` | `social_reformer` | `non_liberal` | `practice` | `international_influence` | `unclassified` - `canon_status`: `core` | `extended` | `referenced` | `unclassified` (editorial centrality on the liberal-canon axis) - `vocations`: array of role values from a closed enum (`philosopher`, `economist`, `statesman`, `industrialist`, `judge`, `scientist`, `writer`, etc.) — see content.config.ts for the full list For primary works: - `title.main`, `title.original_script`, `title.translit`, `title.translation` - `work_type`: `book` | `pamphlet` | `speech` | `essay` | `edited_volume` - `authors`: thinker IDs - `publication.year`, `publication.publisher_name`, `publication.language` - `pdf_url`: R2 URL - `ai_summary`, `ai_key_points` - `paragraph_ids[]`, `clean_markdown_url`: tier-promotion hooks (empty in v1) ## Sibling endpoints - `.md` — markdown body of any HTML page - `/llms.txt` — curated index in the llms.txt spec - `/llms-full.txt` — every Tier-A piece in full plus every Tier-B summary - `/thinkers/.md`, `/primary-works/.md`, etc. ## MCP server A Model Context Protocol server lives at `mcp.indianliberals.in` with the v1 tool surface: | Tool | Scope | Returns | |---|---|---| | `list_works` | A + B | Catalogue, filterable by author, era, theme, kind | | `list_thinkers` | profiles | Every thinker with bio snippet | | `get_work_metadata` | A + B | Author, year, publisher, summary, key points, themes, pdf_url for B, read_url for A | | `read_clean_content` | A only | Body of a musing, opinion, interview, profile, or org page as markdown with `#p-xxxxxx` annotations | | `get_passage` | A only | Specific paragraphs by ID — the citation primitive | | `search_corpus` | All | Pagefind-backed search; results flagged by tier so agent knows whether to quote or summarise | | `find_related` | All | TF-IDF cross-links across content kinds | | `read_index` | — | The curated `/llms.txt` | ## Crawler policy - AI crawlers (ClaudeBot, GPTBot, OAI-SearchBot, PerplexityBot, Google-Extended) are explicitly allowed. - Search engines are blocked from `/theprint-mirror/` so ThePrint keeps the canonical SEO weight. - Editors can toggle individual content via the `noindex` frontmatter field. ## What this archive is NOT - A retrieval chatbot. The host LLM (whichever client you bring) does the reasoning. The MCP server only returns text and structure. - A vector database. Pagefind + structured metadata + `.md` siblings give agents everything similarity search would, plus the structure that similarity search flattens into noise. - A primary-work full-text search (in v1). PDFs are linked, summaries are attributable, paragraph-level citation inside primary works waits for the next engagement. ## Status v1, in active build by Thothica. The site is migrating from a WordPress deployment that ran from 2009 through 2025. Some content is imported and marked `needs_review: true` pending editorial pass; treat the `needs_review` flag as a hint to be careful about citation precision.