Private Subtitle Semantic Search

This is a private, English-only subtitle research tool. It parses authorized .srt or .vtt files locally, stores source cues and normalized gte-small embeddings in a private Supabase project, and returns ranked dialogue with exact timestamps.

Only import subtitles you are authorized to possess and use for personal research. This repository does not provide subtitle files, and its seed data is wholly synthetic.

Requirements

Node.js and npm
Supabase CLI
Docker Desktop for local Supabase testing
A Supabase project you control for hosted use
Optional OpenSubtitles credentials for the download command

The current pipeline accepts English subtitle tracks only. It does not translate, detect, or search non-English subtitle content.

Environment

Create .env from .env.example and set the following values for the project you intend to use:

SUPABASE_URL=https://<project-ref>.supabase.co
SUPABASE_PUBLISHABLE_KEY=<publishable-key>
SUBTITLE_PERSONAL_TOKEN=<long-random-secret>
OPENSUBTITLES_API_KEY=<opensubtitles-api-key>
OPENSUBTITLES_TOKEN=<opensubtitles-user-token>
OPENSUBTITLES_USER_AGENT=private-subtitle-search v1.0

OPENSUBTITLES_API_KEY, OPENSUBTITLES_TOKEN, and OPENSUBTITLES_USER_AGENT are required only when using subtitle download. Obtain and use OpenSubtitles credentials in accordance with its terms and your account permissions. Keep .env, database passwords, and personal tokens out of source control.

Video Candidate Matching Configuration

The Vecteezy matching workflow uses server-only Edge Function secrets. Do not put these values in a browser, CLI request, commit, or client-side application:

VECTEEZY_ACCOUNT=<account-id>
VECTEEZY_API_KEY=<api-key>
AI_INFERENCE_API_HOST=https://<authenticated-ollama-gateway>
OLLAMA_MODEL=gemma4:12b
VIDEO_PLANNER_TRANSPORT=supabase-ai
OLLAMA_GATEWAY_SECURITY_CONFIRMED=true

Before enabling hosted inference, an unauthenticated GET /api/tags must return HTTP 401 or 403 from AI_INFERENCE_API_HOST. A public response means the gateway is not ready: keep hosted inference disabled and fix its authentication. Never bypass this gate for production.

The supabase-ai transport instantiates Supabase.ai.Session('gemma4:12b') directly; this code path does not itself read AI_INFERENCE_API_HOST or OLLAMA_MODEL, and its hosted compatibility with AI_INFERENCE_API_HOST is not yet proven. Keep hosted inference disabled until both the security gate and the hosted compatibility spike pass.

The direct authenticated fallback is available only for a protected Ollama endpoint. The ollama-http transport explicitly consumes AI_INFERENCE_API_HOST, OLLAMA_AUTH_TOKEN, and OLLAMA_MODEL:

VIDEO_PLANNER_TRANSPORT=ollama-http
OLLAMA_AUTH_TOKEN=<bearer-token>

For an internal test gateway with explicit approval for unauthenticated calls, omit OLLAMA_AUTH_TOKEN and set this server-only exception:

VIDEO_PLANNER_TRANSPORT=ollama-http
OLLAMA_ALLOW_UNAUTHENTICATED_TEST_GATEWAY=true

Never enable this test exception in production or for a publicly accessible Ollama service. It only controls whether the direct transport sends a Bearer header; all normal request authentication and response sanitization remain in place.

A local GET /api/tags check is not sufficient for hosted compatibility: verify the gateway from a deployed Edge Function as well. A tunnel or proxy that returns a text/html interstitial to Supabase instead of Ollama JSON will cause the planner to use the deterministic visual-concept fallback.

Local Testing

Local Supabase uses Docker. The seed adds one synthetic ready track with three synthetic cues and two normalized 384-dimensional vectors; it never adds real subtitle dialogue.

npm install
npx supabase start
npx supabase db reset
npx supabase test db
npm test
npm run typecheck

To exercise both Edge Functions locally, set a non-production token in the current PowerShell session and start the functions:

$env:SUBTITLE_PERSONAL_TOKEN='local-test-token'
npx supabase functions serve --no-verify-jwt

In a second PowerShell session, an invalid custom token must receive HTTP 401 from both routes:

Invoke-WebRequest -Method Post -Uri http://127.0.0.1:54321/functions/v1/ingest-subtitles -Headers @{ 'x-subtitle-token' = 'wrong-token' } -ContentType 'application/json' -Body '{"action":"start"}'
Invoke-WebRequest -Method Post -Uri http://127.0.0.1:54321/functions/v1/search-subtitles -Headers @{ 'x-subtitle-token' = 'wrong-token' } -ContentType 'application/json' -Body '{"query":"lantern"}'

Both commands are expected to report a 401 response. Do not use a hosted token for local testing.

Remote Deployment

Remote deployment changes hosted project state. Link the CLI to a Supabase project you control before running the commands below, and set the local CLI environment values required by your project. The Edge Runtime receives SUPABASE_URL and SUPABASE_SERVICE_ROLE_KEY from Supabase; set the separate personal token as a function secret.

npm install
npx supabase login
npx supabase link --project-ref kwoppqigrtvgmmbnzbpx
npx supabase db push
npx supabase secrets set SUBTITLE_PERSONAL_TOKEN=<random-secret>
npx supabase functions deploy ingest-subtitles --no-verify-jwt
npx supabase functions deploy search-subtitles --no-verify-jwt
npx supabase functions deploy movie-quote-montage --no-verify-jwt
npx supabase functions deploy hybrid-subtitle-search --no-verify-jwt
npx supabase functions deploy seed-visual-concepts --no-verify-jwt
npx supabase functions deploy match-video-assets --no-verify-jwt
npx supabase functions deploy select-video-asset --no-verify-jwt
npx tsx src/cli.ts import <authorized-file.srt> --title "The Shawshank Redemption" --year 1994 --imdb tt0111161 --source manual
npx tsx src/cli.ts search "hope during hard times"

Migrations must be pushed before any function is deployed. The project ref above is the target project for this workflow; substitute a different ref only when intentionally deploying elsewhere. Remote deployment compilation is mandatory because local Deno semantic checking is not available in every Node development environment.

Verify the linked migration state and run hosted database advisors after deployment:

npx supabase migration list
npx supabase db lint --linked --level warning

The schema and Edge Functions are deployed to the project above. Deployment does not import subtitle content; import only English subtitle files you are authorized to retain and use.

Video Candidate Matching

match-video-assets turns a stored chunk, supplied text, or a theme into 5-10 Vecteezy video candidates for manual selection. It uses exactly three English search query kinds: literal, action, and metaphor. Candidate counts are in the inclusive range 5-10; use the default of eight unless a review needs a smaller or larger short list.

Push the matching migration before deploying any of its functions. Then set the server-only values above with npx supabase secrets set, deploy in this exact order, and seed the private fallback concepts once:

npx supabase db push
npx supabase functions deploy seed-visual-concepts --no-verify-jwt
npx supabase functions deploy match-video-assets --no-verify-jwt
npx supabase functions deploy select-video-asset --no-verify-jwt

0..23 | ForEach-Object {
  Invoke-RestMethod -Method Post `
    -Uri "https://kwoppqigrtvgmmbnzbpx.supabase.co/functions/v1/seed-visual-concepts?index=$_" `
    -Headers @{ 'x-subtitle-token' = $env:SUBTITLE_PERSONAL_TOKEN } `
    -ContentType 'application/json' `
    -Body '{}'
}

The loop intentionally generates one native gte-small embedding per Edge invocation to stay within the hosted compute budget. Every indexed upsert is idempotent, so rerunning the complete loop is safe.

Use a generic synthetic visual theme or non-sensitive context in requests. Do not send real movie dialogue as ad hoc text. This request returns an idempotent run and temporary previews for review; retrying the same request reuses the persisted result rather than repeating planning and provider search:

$match = Invoke-RestMethod -Method Post `
  -Uri https://kwoppqigrtvgmmbnzbpx.supabase.co/functions/v1/match-video-assets `
  -Headers @{ 'x-subtitle-token' = $env:SUBTITLE_PERSONAL_TOKEN } `
  -ContentType 'application/json' `
  -Body '{"theme":"a fresh start after uncertainty","candidateCount":8}'

Invoke-RestMethod -Method Post `
  -Uri https://kwoppqigrtvgmmbnzbpx.supabase.co/functions/v1/select-video-asset `
  -Headers @{ 'x-subtitle-token' = $env:SUBTITLE_PERSONAL_TOKEN } `
  -ContentType 'application/json' `
  -Body ("{`"runId`":`"{0}`",`"providerResourceId`":{1},`"note`":`"selected after manual review`"}" -f $match.runId, $match.candidates[0].providerResourceId)

Inspect private matching records in the Supabase Dashboard with a service-role administrator only: visual_concepts, video_search_runs, video_search_queries, video_search_candidates, and video_asset_selections. These tables have forced RLS and are not a public API. The selection endpoint records one manual selection for a run and replaces a previous selection atomically.

This phase does not call a Vecteezy download endpoint, does not fetch media, does not store preview URLs, and does not train on Vecteezy material. It only persists stable candidate metadata and the selected provider resource ID.

Rollback and deactivation

To stop the workflow immediately, stop calling the two matching endpoints or delete the match-video-assets and select-video-asset deployments; the subtitle search endpoints remain independent. Use a reviewed follow-up migration to remove the five private matching tables and their RPCs only when their audit records are no longer needed. Do not call Vecteezy media or download routes as part of rollback.

Local Video Workbench

The localhost workbench combines continuous subtitle retrieval, bilingual planning, incremental Vecteezy review, explicit selection confirmation, silent FFmpeg rendering, task history, and recovery in one operational interface. It binds only to 127.0.0.1; provider credentials and temporary provider URLs stay inside the Node control service and are never returned to the browser.

Set the existing Supabase, Vecteezy, and Ollama values plus these local values in .env:

AI_INFERENCE_API_HOST=http://<approved-internal-ollama-host>
OLLAMA_MODEL=qwen3:30b
WORKBENCH_PORT=4173
WORKBENCH_ARTIFACT_ROOT=artifacts
WORKBENCH_FONT_PATH=C:\Windows\Fonts\msyh.ttc
WORKBENCH_FIXTURE_MODE=0

The approved internal Ollama endpoint uses plaintext HTTP and receives the exact ordered English cues selected for the video. The status bar keeps this warning visible. Use this exception only for the explicitly approved test endpoint; use authenticated HTTPS before sending sensitive dialogue anywhere else.

Start the development interface or the built interface:

npm run workbench:dev

npm run workbench:build
npm run workbench:start

The server prints the actual loopback URL and moves to the next available port if the configured port is occupied. Health checks verify Supabase, the selected Ollama model, Vecteezy account access, FFmpeg, ffprobe, the font, and disk space. The Vecteezy health check uses an ordinary search and never calls the formal download endpoint.

Ollama model health verifies availability through /api/tags; it does not guarantee that scene planning will finish inside the 60-second request budget. Workbench generation requests disable Qwen thinking to reduce latency. If task creation times out while health remains available, use a faster model or host and update OLLAMA_MODEL before retrying.

Use the Subtitle library tab to search ready English tracks. English queries go directly to hybrid search; Chinese queries are normalized to a concise English query through the configured Ollama model before search. Choose the result count in the toolbar, then select Create from this quote to return to video production with the clicked movie, subtitle track, and cue midpoint held as the source anchor. The current aspect ratio and 5-10 scene count are reused.

Sync new movies supports two operator-controlled modes. Automatic mode needs an explicit confirmation and continues through the classic candidate list until it reaches the OpenSubtitles quota, exhausts the candidates, or is stopped. Manual mode imports exactly one title after validating a release year and an IMDb ID in tt plus digits format. Stop is cooperative: the current provider or database operation finishes before the job enters stopped. Subtitle synchronization performs real OpenSubtitles downloads and consumes the account's download quota; it does not make a formal Vecteezy download call.

Synchronization state remains local and resumable under:

.batch-state/classic-import-state.json
.batch-state/subtitle-sync-snapshot.json
downloads/classics/

Creating a task selects 5-10 consecutive cues from one ready subtitle track and loads exactly eight review candidates per scene. Load more appends up to eight deduplicated candidates and preserves prior pages and selections. Candidate search, preview, selection, and confirmation consume no formal download quota. Production remains disabled until every scene has one current explicit confirmation; there is no automatic confirmation.

Starting production reserves and performs one formal Vecteezy download for each confirmed scene, so the initial set consumes 5-10 formal calls. A replacement after a deterministic source rejection can use only the task's remaining capacity, and no task can exceed ten formal calls; a 10-scene task therefore has no replacement capacity. Only one production run executes at a time in the local process. Other review tasks remain available in history. A retryable failure can be reopened from history and resumed; uncertain provider reservations fail closed rather than silently spending another formal call.

Version-2 files remain local under:

artifacts/video-runs/<task-id>/manifest-v2.json
artifacts/video-runs/<task-id>/review-state.json
artifacts/video-runs/<task-id>/sources/
artifacts/video-runs/<task-id>/subtitles.ass
artifacts/video-runs/<task-id>/final.mp4

The final MP4 is H.264, yuv420p, 30 fps, and has no audio stream. The workflow does not add music, ambience, narration, TTS, or a Storage upload.

Before using the real workbench, apply the additive v2 migration and deploy the three changed Edge Functions in this order:

npx supabase db push --linked
npx supabase functions deploy subtitle-passages --no-verify-jwt
npx supabase functions deploy match-video-assets --no-verify-jwt
npx supabase functions deploy video-production-metadata --no-verify-jwt

The v2 schema, RPCs, API actions, manifests, and renderer are additive. Existing v1 video plan, video produce, and video resume commands and completed v1 manifests remain unchanged.

For browser verification, install Chromium once and run the test-only fixture:

npx playwright install chromium
npm run test:e2e

WORKBENCH_FIXTURE_MODE=1 is accepted only with NODE_ENV=test; production startup rejects that combination before reading credentials. The fixture makes no Supabase, Ollama, OpenSubtitles, or Vecteezy request and therefore consumes neither OpenSubtitles quota nor formal Vecteezy downloads.

Local Video Production

This resumable workflow keeps every source clip, review file, subtitle file, render, contact sheet, manifest, and retry record under ignored artifacts/. There is no Storage upload: media remains on the trusted operator machine. The three private metadata tables are video_render_jobs, video_asset_downloads, and video_render_segments.

Remote-only migration and database testing are required for this workflow. Review the linked project before applying changes, then run the dry run, migration, focused database test, and metadata deployment in this order:

npx supabase db push --dry-run --linked
npx supabase db push --linked
npx supabase test db --linked supabase/tests/database/video_production.sql
npx supabase db query --linked --file supabase/tests/database/video_production.sql
npx supabase functions deploy video-production-metadata --no-verify-jwt

The db query form is the Docker-free fallback for CLI builds where linked test db still invokes a local container. The test transaction temporarily enables pgTAP and rolls back the extension together with all fixture rows.

Local rendering requires ffmpeg and ffprobe on PATH, plus the Microsoft YaHei font at the renderer's configured Windows font path. Set the existing Supabase, subtitle token, Vecteezy, and planner environment values in .env. The plan command requires only SUPABASE_URL, SUPABASE_PUBLISHABLE_KEY, and SUBTITLE_PERSONAL_TOKEN; produce and resume also require VECTEEZY_ACCOUNT and VECTEEZY_API_KEY.

Create a plan with the approved generic theme:

$plan = npm run --silent video -- plan --theme "Crossing darkness toward dawn" --candidate-count 8 --json | ConvertFrom-Json

The command emits one path-only JSON object and writes the same pointer to artifacts/latest-plan.json. Planning searches the private subtitle index, protects the exact quote so it is never trimmed or rewritten, and creates four generic visual searches. It sends no dialogue to Ollama. Plan does not consume any Vecteezy download quota and makes no formal download request.

Open the generated review-candidates.json beside $plan.reviewPath, inspect its temporary previews locally, and fill the strict review-input.json with a reviewed Chinese quote translation, one listed resource ID per scene, a review note, and a nonnegative source in-point. Review must finish before produce. Do not paste signed or status URLs into the input.

npm run video -- produce --manifest $plan.manifestPath --review $plan.reviewPath --max-downloads 4

The production contract has a hard exactly-four-download budget. The runner records four manual selections and starts one idempotent render job to learn the existing status before any download-info preflight. An existing completed job short-circuits with zero Vecteezy provider calls. For a fresh job, download-info preflight validates sizes and attribution before local render ownership or any formal provider call. A fresh preflight rejection restores the local plan to editable review so the operator can replace the candidate and rerun produce; the already-created remote job may remain failed as an audit record. Only after preflight succeeds does the runner permit at most four formal provider calls. Matching local hashes are reused. An uncertain transfer reservation fails closed and requires a new reviewed plan, not resume, so the run can never silently spend the reservation again. A render failure preserves verified source files for resume, and a completed local manifest is immutable even when remote completion needs another attempt.

In a later PowerShell session, resume from the current render-owned path rather than guessing a UUID:

$active = Get-Content -Raw artifacts/latest-plan.json | ConvertFrom-Json
npm run video -- resume --manifest $active.manifestPath

Rollback and deactivation

Stop video produce and video resume, then delete or deactivate the video-production-metadata function deployment. Existing local files remain under ignored artifacts/ and are not uploaded. Remove video_render_segments, video_asset_downloads, and video_render_jobs only through a reviewed follow-up migration after their audit records are no longer required. The subtitle search and candidate matching workflows remain independent.

Movie Quote Montage

movie-quote-montage accepts an English theme and returns a deterministic montage of exact stored subtitle chunks. It limits how many chunks can come from one movie, so a single title does not dominate the result. The copy field joins the selected chunks with blank lines; it is retrieved dialogue, not newly generated prose.

The Edge Function generates the query vector inside Supabase with new Supabase.ai.Session('gte-small'), using the same normalized 384-dimensional model as the stored chunks. No external embedding API is required.

After pushing the migration and deploying the function, test a theme such as love and time with the existing private token:

Invoke-RestMethod -Method Post `
  -Uri https://kwoppqigrtvgmmbnzbpx.supabase.co/functions/v1/movie-quote-montage `
  -Headers @{ 'x-subtitle-token' = $env:SUBTITLE_PERSONAL_TOKEN } `
  -ContentType 'application/json' `
  -Body '{"theme":"love and time","quoteCount":8,"matchThreshold":0.72,"maxPerMovie":1}'

The response includes copy plus source metadata for every selected chunk: movie title, release year, timestamps, chunk indexes, and similarity score. Raise matchThreshold for stricter matches or lower it when the result set is empty.

Hybrid RRF Experiment

hybrid-subtitle-search is an isolated all-Supabase experiment based on the official Supabase hybrid-search pattern. PostgreSQL retrieves one candidate list with English full-text search and another with the existing gte-small cosine vectors, then combines their ranks with Reciprocal Rank Fusion (RRF). The existing search-subtitles endpoint remains unchanged as the vector-only baseline.

The initial experiment fixes these RPC controls:

full_text_weight = 1
semantic_weight = 2
rrf_k = 50

The higher semantic weight prevents short literal phrases from overwhelming strong thematic matches. These values are experimental and should be changed only after comparing returned dialogue.

After pushing the migration and deploying hybrid-subtitle-search, call both endpoints with the same private token and request body:

$body = '{"query":"love and time","limit":12}'
$vector = Invoke-RestMethod -Method Post `
  -Uri https://kwoppqigrtvgmmbnzbpx.supabase.co/functions/v1/search-subtitles `
  -Headers @{ 'x-subtitle-token' = $env:SUBTITLE_PERSONAL_TOKEN } `
  -ContentType 'application/json' `
  -Body $body
$hybrid = Invoke-RestMethod -Method Post `
  -Uri https://kwoppqigrtvgmmbnzbpx.supabase.co/functions/v1/hybrid-subtitle-search `
  -Headers @{ 'x-subtitle-token' = $env:SUBTITLE_PERSONAL_TOKEN } `
  -ContentType 'application/json' `
  -Body $body

Hybrid results include the normal similarity, source, timestamps, text, and cues plus three diagnostics. rrfScore is the final fused score; semanticRank is the candidate's vector rank; and fullTextRank is its keyword rank. Either rank can be null when the candidate appeared in only one of the two bounded candidate lists.

The experiment is independently reversible. Stop calling or delete the hybrid-subtitle-search deployment, then use a reviewed follow-up migration to drop public.hybrid_match_subtitle_chunks, subtitle_chunks_fts_gin_idx, and public.subtitle_chunks.fts. Do not run ad-hoc destructive SQL as part of the normal deployment workflow. search-subtitles remains available throughout.

The CLI can also retrieve a subtitle through the official OpenSubtitles API when your credentials and rights permit it:

npx tsx src/cli.ts download --imdb 0111161 --output downloads/authorized.srt
npx tsx src/cli.ts import downloads/authorized.srt --title "Authorized Title" --year 2026 --imdb tt0000001 --source opensubtitles --source-ref opensubtitles:42
npx tsx src/cli.ts search "quiet determination" --limit 10
npx tsx src/cli.ts search "quiet determination" --movie-id 7

To add a large classic-film batch, use the resumable importer. It uses the ranked candidate pool in data/classic-movie-candidates.json, stores downloaded subtitle files under ignored downloads/classics, and writes resume progress to ignored .batch-state/classic-import-state.json.

npx tsx src/batch-classics.ts --target 200

If OpenSubtitles returns a daily quota or rate-limit response, the command stops cleanly. Rerun the same command later to continue from the saved state.

Search output is compact and timestamped:

0.842  00:42:13.120 --> 00:42:18.900
The matching dialogue text...

Privacy And Data Lifecycle

All subtitle tables use forced row-level security with no policies for anon or authenticated. Direct table, sequence, and match_subtitle_chunks access is revoked from public roles; only the Edge Functions' service role can read and write subtitle data after custom-token authentication. Search fetches only the exact cue interval for each ranked chunk, never an entire track.

Back up only data you are entitled to retain. For a hosted project, an operator can export the private tables with a protected database URL:

pg_dump --data-only --table=public.movies --table=public.subtitle_tracks --table=public.subtitle_cues --table=public.subtitle_chunks "$env:DATABASE_URL" > subtitle-backup.sql

Deleting a movie cascades to its tracks, cues, embeddings, and chunk claims:

delete from public.movies where imdb_id = 'tt0111161';

Use the movie ID or a more specific predicate when removing content. Verify the selected rows before running destructive SQL.

Troubleshooting

Docker is unavailable: npx supabase start, db reset, test db, and local function serving require Docker Desktop with a running engine. Start Docker Desktop, verify docker version, then rerun the local workflow.
Function secret missing: set SUBTITLE_PERSONAL_TOKEN with npx supabase secrets set SUBTITLE_PERSONAL_TOKEN=<random-secret>, redeploy both functions, and use that same value in the CLI .env file.
HTTP 401: confirm the request sends x-subtitle-token, the value exactly matches the function secret, and the function deployment is the intended project. The Supabase publishable key does not replace this custom token.
OpenSubtitles HTTP 429: wait for the Retry-After duration and retry within your account quota. The client retries bounded transient responses, but it cannot override provider limits.
Model dimension error: this project requires gte-small embeddings with mean_pool: true, normalize: true, and exactly 384 finite numeric dimensions. Do not mix vectors from another model or change the vector column dimension.
No matching dialogue found: this is a successful search with no ranked results. An empty_ready_track API error means the requested movie has no finalized ready subtitle track; import authorized English subtitles and finalize the track first.

Name		Name	Last commit message	Last commit date
Latest commit History 146 Commits
.superpowers/sdd		.superpowers/sdd
data		data
docs/superpowers		docs/superpowers
e2e		e2e
scripts		scripts
src		src
supabase		supabase
test-results		test-results
tests		tests
workbench		workbench
.env.example		.env.example
.gitignore		.gitignore
README.md		README.md
package-lock.json		package-lock.json
package.json		package.json
playwright.config.ts		playwright.config.ts
tsconfig.json		tsconfig.json
vite.config.ts		vite.config.ts

Navigation Menu

Repository files navigation

Private Subtitle Semantic Search

Requirements

Environment

Video Candidate Matching Configuration

Local Testing

Remote Deployment

Video Candidate Matching

Rollback and deactivation

Local Video Workbench

Local Video Production

Rollback and deactivation

Movie Quote Montage

Hybrid RRF Experiment

Privacy And Data Lifecycle

Troubleshooting

About

Releases

Packages

Contributors

Languages

Folders and files

Latest commit

History

Repository files navigation

Private Subtitle Semantic Search

Requirements

Environment

Video Candidate Matching Configuration

Local Testing

Remote Deployment

Video Candidate Matching

Rollback and deactivation

Local Video Workbench

Local Video Production

Rollback and deactivation

Movie Quote Montage

Hybrid RRF Experiment

Privacy And Data Lifecycle

Troubleshooting

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages