Mimir

Find what to watch, read, play or listen to next. Mimir is a free catalogue and discovery engine for films, TV series, books and manga, games, music and YouTube, built by Terratales Media Entertainment. Live at mimir.terratalesmedia.world.

Describe what you want to experience: a mood, an idea, a favourite work, a time limit. Mimir returns a few well-chosen picks, each with the reasons it fits. The catalogue holds about 11,000 works from 101 countries in 68 original languages; 681 are fully classified against 372 controlled terms (233 of them subgenres), and the rest, imported from Wikidata, are being classified in batches.

Under the hood, a request in plain language is interpreted into inspectable constraints and preferences, a classified catalogue is ranked against them, and every result is explained from recorded evidence. Facts are labelled separately from interpretations. There is no LLM and no API key.

It is also a browsable catalogue, in the spirit of IMDb, MyAnimeList and Goodreads in one place: directories per medium, 36 shelves, subgenres, tag and people pages, a parents' guide for every medium, "where to watch" links, a personal library with activity rings and badges, shareable lists, watch-alongs and a community.

Ways to tell Mimir your taste, besides rating works: swipe through a deck (right like, up love, left not for me), spin a mood when you have no time to browse, craft hybrids by combining works, moods and genres, and open Days of the year (World Laughter Day, International Music Day and about 50 more) for something new and picks from your own profile. Franchise pages gather a whole IP (James Bond, Middle-earth, the Marvel Cinematic Universe, the Wizarding World) with a row per medium, and films and series list their lead cast. One click in your profile resets your recommendations.

• Everything Mimir does: docs/FEATURES.md
• Press kit and press release: docs/press/
• Search visibility: docs/SEO-KEYWORD-STRATEGY.md (keywords, meta map, what is built and what is not).
• How the catalogue grew to about 11,000 works, and what is still being filled in: docs/CATALOGUE-EXPANSION.md · Planned features: docs/FUTURE-FEATURES.md
• Help centre: the site's own guides, FAQs and walkthrough video at mimir.terratalesmedia.world/#help
• Beta testers start here: docs/BETA.md
• Where Mimir stands against MAL, MangaUpdates, IMDb, Letterboxd, Goodreads and Spotify: docs/COMPETITIVE.md · What comes next: docs/ROADMAP.md
• Brand guide (name, mark, colour, type, voice):

Layout

api/                    Python 3.11+, FastAPI, PostgreSQL (modular monolith)
  mde/migrations/       SQL migrations, applied in order, additive only
  mde/taxonomy/         Controlled taxonomy (taxonomy.json) + loader/validator
  mde/seed/             Editorial seed: catalogue.json (the pilot, with defaults) plus films, series, books, games,
                        albums, food, YouTube and more (*.json), merged by catalogue/seedfiles.py
  mde/data/enrichment/  Open-data enrichment written by `python -m mde ingest` (Wikidata, Open Library, MusicBrainz)
  mde/ingest/           Connectors, matching rules and the HTTP cache
  mde/catalogue/        Import (load.py) and read-side queries (repository.py)
  mde/discovery/        intent.py → rank.py → explain.py, orchestrated by engine.py
  mde/app.py            HTTP API
  benchmark/            Discovery benchmark (queries.json) and runner
  tests/                Unit tests + integration tests against a throwaway database
web/                    Next.js 15 + TypeScript (server components; no client-side API calls)
brand/                  Mimir mark, lockups, icons and social card (see docs/BRAND.md)
demo/                   The public single-page site: a JavaScript port of the pure pipeline (engine.cjs),
                        kept identical to the Python by api/tests/test_demo_parity.py

Public site

node demo/build.cjs --site writes demo/dist/site/, the deployable static site for mimir.terratalesmedia.world, with a full <head>, icons, social card and CNAME. .github/workflows/pages.yml publishes it on every push to main (setup in docs/DEPLOY.md). node demo/build.cjs alone writes a single self-contained page that runs the discovery pipeline (intent parsing, ranking, explanations) on the pilot catalogue with no server. Differences from the full stack: keyword search is a simple word match instead of PostgreSQL full-text search, and feedback stays in the viewer's browser. pytest runs the JavaScript and Python engines side by side on every benchmark query plus about 40 edge cases, and fails if their output differs in any field.

Running locally

Requirements: Python 3.11+, Node 22, PostgreSQL 16.

# 1. Database
createdb mde                       # or any database; set DATABASE_URL accordingly

# 2. API
cd api
python -m venv .venv && . .venv/bin/activate
pip install -e ".[dev]"
export DATABASE_URL=postgresql://postgres:postgres@localhost:5432/mde
python -m mde setup                # apply migrations + import the pilot catalogue (safe to re-run)
uvicorn mde.app:app --port 8000

# 3. Web (another terminal)
cd web
npm ci
API_URL=http://localhost:8000 npm run dev     # http://localhost:3000

Try it from the command line too:

python -m mde discover "a book like Outer Wilds"
python -m mde benchmark

Environment variables

Variable	Used by	Default	Purpose
DATABASE_URL	api	postgresql://postgres:postgres@localhost:5432/mde	Catalogue database
CORS_ORIGINS	api	http://localhost:3000	Comma-separated browser origins (the web app calls the API server-side, so this only matters for direct browser use)
API_URL	web	http://localhost:8000	Where the web server reaches the API
TEST_DATABASE_ADMIN_URL	tests	postgresql://postgres:postgres@localhost:5432/postgres	Tests create and drop their own database through this connection

No API keys are needed. Discovery uses no external providers and no LLM.

Checks

cd api && ruff check . && pytest -q          # DB tests skip without PostgreSQL, parity tests without Node
cd web && npm run typecheck && npm run build

CI (.github/workflows/ci.yml) runs the same checks against a PostgreSQL 16 service.

API

Method	Path
GET	/api/discover?q=…&limit=8	Interpretation, ranked results with reasons, caveats and score breakdowns, and the exclusion counts. Logs an anonymous discovery event
GET	/api/works?medium=&tag=&avoid=&decade=&language=&country=&max_minutes=&q=&sort=	Browse: filtered, sorted work cards plus facet counts. tag and avoid take taxonomy keys and repeat; avoid keeps works assessed as none or mild
GET	/api/works/{slug}	Work page: facts, credits, tags by dimension, parents' guide (official ratings plus per-category severity and notes), connected works, provenance, sources
GET	/api/tags, /api/tags/{dimension}/{slug}	Tag index with counts; a tag's definition, its works (strongest first) and related tags
GET	/api/people, /api/people/{slug}	Everyone credited; a person's or studio's works, roles and recurring qualities
GET	/api/works/{slug}/where?region=IN	Where to watch, listen, read or play: a link per service for that region (search links unless the catalogue holds an exact ID)
GET	/api/shelves, /api/shelves/{slug}	Wordplay shelves (Laugh Riot, Drama Mamas, Nepo Babies, Desi Delights …)
GET	/api/groups, /api/groups/{slug}	Series and franchise families, in order, with total time
GET	/api/music/moods, /api/music/remixes	Mood rings and the remix rack
POST	/api/recommend	{ratings: {slug: loved|liked|disliked}, limit, medium} → taste profile and picks, each with the work it resembles and the qualities it shares
GET	/api/taxonomy	The controlled vocabulary
POST	/api/feedback	{event_id, work_slug, kind} where kind is one of relevant, irrelevant, too_long, wrong_mood, , ,

Accounts are an encrypted vault keyed by a recovery key, stored by the Mimir Vault WordPress plugin (wordpress/mimir-vault/, see docs/ACCOUNTS.md); the browser code is demo/vault.cjs.

Community (mde/community.py; clients send Authorization: Bearer <token>):

Method	Path
POST	/api/auth/signup, /api/auth/login, /api/auth/logout	Accounts. Passwords hashed with scrypt; only a SHA-256 of each session token is stored; sign-in is rate-limited
GET, PATCH	/api/me	Profile, privacy (private profile, share activity)
POST	/api/me/delete	Deletes the account and everything it wrote
GET, PUT	/api/me/state	The private library (statuses, scores, progress, lists, journal, challenges), synced as one document
GET	/api/users/{username}	Profile: counts, recent reviews, public lists, activity (followers only when private)
POST, DELETE	/api/users/{username}/follow	Follow and unfollow
GET, PUT, DELETE	/api/works/{slug}/reviews, /api/works/{slug}/review	Reviews with a community score; one review per person per work
POST, DELETE	/api/reviews/{id}/like; GET, POST /api/reviews/{id}/comments; DELETE /api/comments/{id}	Likes and comments
POST	/api/activity	"started" / "finished", shared with followers when the user allows
GET	/api/feed, /api/community	Your feed; trending works, latest reviews and activity, challenge counts
POST, GET, DELETE	/api/lists, /api/lists/{id}	Published lists
GET, POST, DELETE, PUT	/api/challenges/{slug}/{year} (/join, /progress)	Social-calendar challenges: members, leaders, your ticked days
POST	/api/reports	Report a review, comment, user or list

Known limitations

• Facts are partly verified. Where Wikidata, Open Library or MusicBrainz match a work, their year, length, page count, scores, box office, covers and track lists are used and the page says which source each came from (and what the seed said). Everything else was compiled from general knowledge and is unverified. Every tag, summary and parents' guide level is an AI-assisted reading awaiting editorial review; the UI says so.
• Episode counts for running TV series are as compiled and will drift; they are not yet fetched.
• "Where to watch" links open each service's search. Mimir cannot confirm a title is on a service in your region; that needs a licensed availability feed (e.g. a JustWatch partnership).
• Posters for films, TV and games need a TMDB or IGDB agreement; until then those works use generated covers. Book and album covers come from Open Library and the Cover Art Archive.
• The benchmark (20 queries) was written alongside the system. It catches regressions but doesn't prove relevance. The beta's job is to gather real judgements.
• The intent parser is rule-based. It covers the phrasing in the benchmark and tests and will miss some phrasings. An LLM-backed parser can implement the same IntentParser interface.
• Ranking, browse and recommendations run in memory over the whole catalogue: fine for thousands of works; retrieval needs to move into SQL/pgvector before tens of thousands.
• Ratings, shelves and feedback on the public site stay in the viewer's browser unless the API is hosted and the viewer signs in (set MIMIR_API_URL for the Pages build; see docs/DEPLOY.md).

Licence

Copyright © 2026 Terratales Media Entertainment Private Limited. All rights reserved; see LICENSE. Third-party fonts and open data keep their own licences, listed in NOTICE.md.