savethenurse/ai-media-hub

Fork 0

Files

T

AI Assistant 75f1bb360c

build-push / docker (push) Successful in 4m8s

Details

Reduce noisy Gemini partial failure counts

2026-03-17 15:09:23 +09:00

54 KiB

Raw Blame History

AI Media Hub Handover

Working Rule

This file is both backlog and handover log.
Every meaningful change should record:
- what changed
- why it changed
- how it was verified
- what is still risky or incomplete
If a push fails or a change remains local-only, that must be written here explicitly.

Current State At A Glance

Project: ai-media-hub
Goal: AI-assisted media discovery + ingest dashboard for Unraid
Backend: Go
Worker: Python + yt-dlp + ffmpeg
Frontend: HTML + Vanilla JS + Tailwind CDN
Database: SQLite
Search backend: SearXNG
AI translation / visual ranking: Gemini 2.5 Flash
Deployment target: single Docker container on Unraid
Git remote: https://git.savethenurse.com/savethenurse/ai-media-hub.git

Current Status Summary

Upload / direct download flow is implemented and broadly usable.
Search is implemented end-to-end and now refactored into source-specific collectors.
Search remains the main unstable subsystem.
Envato metadata and preview extraction are much stronger than before, including additional hydration-data preview fallback.
Artgrid metadata fidelity is improved, but stable public hover-video preview extraction is still not solved.
Frontend now logs more useful API and debug information than earlier versions.
A local self-test workflow now exists and should be run before container builds or pushes.
A fresh-machine bootstrap was revalidated in a user-local toolchain setup on 2026-03-17; go test ./... and bash scripts/selftest.sh now pass in that setup.
Result modal sizing is now being constrained to the viewport, and modal-only source-summary translation is now part of the active implementation path.
Card summaries now also translate lazily to Korean, and Gemini negative-assessment handling now drives stronger follow-up search behavior than before.
Search preview delivery is now moving away from persistent on-disk preview caching toward live proxy / live transcode behavior, with Google Video preview reuse added to result cards and modal playback.
Search breadth and Gemini review budget were widened again because the latest user feedback still reported a thin visible result count even after smarter filtering.
The codebase is now back on the broader-search / modal-fitting direction associated with 5ca7aef, with an added Gemini deadline guard to reduce reverse-proxy 504 risk.

Current Architecture

backend/main.go
- app bootstrap
- env loading
- static frontend serving
- route registration
backend/handlers/api.go
- upload / download / search APIs
- WebSocket progress broadcast
- debug event broadcast
- search request orchestration only, with ranking/Gemini logic mostly moved out
backend/services/cse.go
- SearXNG querying
- shared search helpers
- source-specific enrich helpers
- URL filtering / parsing utilities
backend/services/search_collectors.go
- source-specific collectors:
  - envatoCollector
  - artgridCollector
  - googleVideoCollector
backend/services/ranker.go
- ranking
- Gemini candidate cap logic
- Gemini batch evaluation wrapper
- recommendation merge logic
backend/services/gemini.go
- query translation
- deterministic query expansion
- Gemini vision scoring
- video frame extraction via ffmpeg when needed
backend/models/db.go
- SQLite init
- download history
worker/downloader.py
- yt-dlp probe / download
- ffmpeg clip extraction
frontend/index.html
- main dashboard UI
- result viewer modal
- preview modal
- debug log panel
frontend/app.js
- API calls
- WebSocket status bar
- result viewer modal
- hover preview playback
- direct download handoff for Google Video results
- debug logger panel
- platform toggles
frontend/style.css
- custom styles
- clamp helpers
- slider thumb styles
- debug panel scrollbar styles
scripts/selftest.sh
- local smoke test flow
scripts/mock_searxng.py
- local mock SearXNG used by self-test
unraid-template.xml
- Unraid template for current image source

Search Flow: Current Implementation

User enters a query in Zone A.
Frontend sends /api/search with:
- query
- selected platforms
Backend translates the query in GeminiService.TranslateQuery.
- Gemini translation if available
- Google Translate HTTP fallback
- Korean media-term dictionary fallback
- explicit normalization for known compound phrases such as 사이버 펑크 -> cyberpunk
Backend builds deterministic English search variants in GeminiService.ExpandQuery.
SearchService.SearchMedia(...) orchestrates source-specific collectors.
Collectors query SearXNG separately for:
- Envato
- Artgrid
- Google Video
Each collector applies source-specific acceptance logic.
- Google Video: YouTube-only plus noise filtering
- Envato: elements.envato.com item URLs only
- Artgrid: accepts both:
  - artgrid.io/clip/...
  - artlist.io/stock-footage/clip/...
Artgrid canonical links are normalized to:
- https://artgrid.io/clip/<id>/<slug>
Results are enriched source-by-source.
- Envato:
  - VideoObject JSON-LD preferred
  - page meta preferred over search-engine proxy thumbnail
  - preview mp4 extraction via JSON-LD / HTML parsing
  - Python HTML fetch fallback used when Go HTTP fetch gets Cloudflare challenge pages
- Artgrid:
  - page title / description / thumbnail cleaning
  - homepage / challenge HTML is now rejected so generic site metadata does not overwrite clip metadata
  - preview video extraction still not stable
Ranked results are passed through the shared ranker.
All ranked candidates are evaluated with Gemini Vision in batches.
Merge order now prefers:

Gemini recommended items
Gemini-reviewed non-recommended items
keyword fallback items only if Gemini output is incomplete

Frontend renders cards, result viewer modal, and hover previews.

Direct Downloader Flow: Current Implementation

User enters URL in Zone C.
Frontend checks duplicate history via /api/history/check.
Frontend loads preview metadata via /api/download/preview.
Preview modal opens with:
- media preview
- duration
- crop dual-thumb slider
- quality select
User confirms download.
Backend launches Python worker.
Worker downloads source with yt-dlp, clips with ffmpeg, emits JSON progress lines.
Backend rebroadcasts progress over WebSocket.

Major Work Completed So Far

Added local self-test workflow:
- scripts/selftest.sh
- scripts/mock_searxng.py
Fixed translation fallback when Gemini key is missing.
Added tests for translation fallback logic.
Added HLS frontend wiring:
- hls.js script
- native HLS fallback
Reduced search timeout risk by:
- limiting collector result caps
- limiting enrichment scope
- limiting Gemini Vision evaluation scope
- replacing oversized raw debug result payloads with summaries
Improved Google Video filtering:
- rejects more music / trailer / BGM style noise
Improved Envato fidelity:
- real title / description / thumbnail / preview from source page
Improved Artgrid fidelity:
- accepts canonical Artlist URLs
- normalizes Artgrid clip URLs
- cleans title / description better
Refactored search into source-specific collectors.
Moved ranking and Gemini batch handling into backend/services/ranker.go.
Fixed server-side 500 caused by Gemini candidate cap exceeding available ranked candidates.
Improved frontend logging:
- raw non-JSON error body logging
- more compact debug payload rendering
Changed hover preview playback to lazy attach on hover:
- attach source on mouseenter
- wait for readiness before play()
- detach source on mouseleave
Added in-app result viewer modal for search results:
- results now open in a modal instead of directly opening a new tab
- modal now prefers internal preview media over embedded third-party iframes to avoid embed blocking
- external open button remains available
Google Video results can now jump directly into the existing direct-download preview / crop flow from the result viewer
Gemini reason generation is now intended to be Korean-first for readability
Gemini Vision evaluation now covers all ranked results instead of only a top subset
Search results now prioritize AI note text visually ahead of source summary
Search query order and final top results now include light randomness so repeated searches are less static

Current Features Implemented

Project folder structure
Dockerfile
Gitea workflow
Unraid template
SQLite download history
File upload
yt-dlp direct downloader
Preview modal for direct download
Crop selection slider
Quality selection
WebSocket realtime progress
Search source toggles
Search card hover preview support
Result viewer modal for search results
Google Video direct-download handoff from search results
Debug log panel in frontend
.log download from debug panel
Local self-test workflow
Source-specific search collectors
Shared ranker service layer

Important Current Constraints / Known Problems

Search backend quality is still the most fragile subsystem.
Search relevance is still heuristic-heavy and not yet benchmarked against a durable real-query set.
Embedded third-party result viewing is no longer relied on because many providers block iframe embedding with X-Frame-Options / CSP.
Artgrid hover-video preview is still partial / unresolved:
- provided Artgrid HTML snapshots and downloaded asset bundles did not expose a stable public preview mp4/m3u8 URL
- public HTML often only exposes title / description / thumbnail / canonical URL
Artgrid can still be sensitive to how SearXNG indexes canonical domains.
Full browser-level validation is still not covered by local self-test.
Frontend JavaScript still has no Node-based lint/build step in this environment.
Search cards now separate source snippet from AI reason, but metadata fidelity still depends on source enrichment quality.
Gemini notes are now intended to be Korean, but final output quality still depends on Gemini response consistency.
Source Summary translation now depends on Google Translate HTTP availability; frontend silently falls back to original summary text if translation fails.
The result modal should now stay within viewport height, but this still needs real browser confirmation on multiple short-height displays because CSS-only constraints were the source of the latest user-visible regression.
Artgrid preview playback now has a server-side ffmpeg transcode path for .m3u8 style preview URLs, but this trades storage savings for runtime CPU cost.
The provided Artgrid HTML sample still does not expose a direct preview m3u8 or mp4 URL by itself, and yt-dlp probe on the sample clip URL returned Unsupported URL, so fully reliable Artgrid playback still depends on live-page/runtime preview discovery succeeding elsewhere in the pipeline.
Docker CLI is not installed in this environment, so container-image build verification still cannot be performed locally from this machine.
Gemini batch warnings can still mix true model/output failures with harmless candidate-skip cases unless the recovery path filters those error classes before surfacing them.
The local self-test script is better than before, but it is still a smoke test, not full integration coverage.

Current Risks Around Search Quality

Upstream SearXNG quality still controls the candidate pool.
Gemini Vision can only rerank the candidates it receives.
If source enrichment fails, Gemini may still judge a weaker proxy thumbnail or fallback image.
Compound Korean intents are better handled now, but the translation path is still heuristic and can drift on niche concepts.
Running Gemini Vision across all ranked results increases latency and token usage compared with the earlier capped approach.

Frontend Debug Logger

UI button: bottom-right Logs
Files:
- frontend/index.html
- frontend/app.js
- frontend/style.css
Logs currently capture:
- API request / response
- WebSocket progress messages
- ignored WS debug messages
- status updates
- platform toggle state
- result viewer modal open / close
- preview source attach / detach
- hover start / hover end
- hover play errors
- modal preview open / close
- browser errors
- promise rejections
- backend debug broadcasts

Recent Change Log

Date: 2026-03-16
What changed:
- Stabilized the Gemini visual-review path after widened search budgets caused full-batch “no candidate thumbnails or preview frames” failures.
- Google Video enrichment now always prefers the canonical YouTube ytimg thumbnail instead of keeping a potentially broken search-engine thumbnail.
- Gemini visual fetch now preserves the last concrete fetch error, retries with derived YouTube thumbnails when possible, and only falls back to the generic no-visual message after all image/frame paths fail.
- User-facing fallback warning text for the all-batches-no-visual case is now softened so ranked results can still be shown without surfacing the raw internal Gemini error string in the UI.
Why it changed:
- The latest deployed build widened search enough that top-ranked candidates sometimes carried metadata without any fetchable image bytes, causing Gemini review to fail for every batch and surfacing an alarming warning even though ranked fallback results still existed.
How it was verified:
- go test ./...
- bash scripts/selftest.sh
What is still risky or incomplete:
- This improves YouTube-backed candidates immediately, but Envato/Artgrid thumbnails can still be inaccessible in some provider-side cases, so Gemini can still fall back to ranked results when source media is locked down.
Date: 2026-03-16
What changed:
- Hid the expanded query-variant chip list from the search UI while leaving backend/debug query visibility intact.
- Reworked the result modal action panel so its primary CTA now follows source capability: Direct Download for Google Video and Open Source for Envato/Artgrid.
- Added a secondary action button area plus internal scrolling for long Source Summary text so long descriptions no longer push action buttons off-screen.
- Switched Google Video modal rendering from iframe-style embed-first behavior to a webpage-like preview panel, while keeping Artgrid/Envato on the existing media-mode path.
- Bumped the frontend asset version again to force browsers onto the updated modal behavior.
Why it changed:
- The remaining user issues were now primarily UX/layout issues: too little visible result choice, hidden buttons under long summaries, query-variant clutter, and modal actions that did not match what each source can actually do.
How it was verified:
- go test ./...
- bash scripts/selftest.sh
What is still risky or incomplete:
- Live browser validation is still needed to confirm the widened search pool feels better in real queries and that the new Google Video panel feels sufficiently “webpage-like” in practice.
- node is still unavailable in this environment, so no frontend static syntax/build step was added here.
Date: 2026-03-16
What changed:
- Expanded search breadth moderately by increasing base query count, collector query budgets, per-source caps, enrichment scope, and final visible result target while keeping Gemini review cap at 16.
- Reworked recommendation action metadata so Google Video now advertises Direct Download as the primary CTA, while Envato and Artgrid advertise Open Source.
- Changed default modal media priority so Artgrid now prefers preview video ahead of thumbnail when both are available, and Google Video now defaults to a webpage-like thumbnail mode instead of embed-first.
- Added visible-count style debug summary fields to support checking whether the widened search budget actually increases user-facing choice.
Why it changed:
- The UI was much healthier, but the remaining request from the user was to widen the pool of selectable results without undoing the recent quality gains, and to align modal CTA semantics with what each source can actually do.
How it was verified:
- go test ./...
What is still risky or incomplete:
- The widened search budget may increase latency on worse SearXNG days, so the frontend/UI half of this batch still needs to land before the full user-facing behavior is validated.
Date: 2026-03-16
What changed:
- Rewired the result modal to consume backend media metadata instead of hard-coded source branches.
- Google Video now uses embed-first modal rendering again, with iframe timeout fallback to thumbnail/panel mode.
- Search cards now suppress low-value favicon/logo thumbnails and show a neutral “preview unavailable” media state instead of tiny site icons or placeholder-like junk.
- Bumped frontend asset version so browsers pick up the new modal logic.
Why it changed:
- The UI was still rendering Google Video as a static image panel, and Envato/Artgrid cards were surfacing unusable thumbnails that made the results look broken even when metadata existed.
How it was verified:
- go test ./...
- bash scripts/selftest.sh
What is still risky or incomplete:
- Browser-level validation was still not fully reproducible here, so iframe timeout behavior and provider-specific rendering quirks still need live confirmation in the deployed UI.
- node is not installed in this environment, so frontend syntax/build verification is still limited to static inspection plus app boot smoke testing.
Date: 2026-03-16
What changed:
- Hardened search result enrichment and recommendation metadata for preview recovery work.
- Added provider-aware fetch strategy for source HTML/JSON requests, broader Envato preview parsing, looser Artgrid HTML acceptance, and stronger thumbnail preservation rules.
- Added low-value thumbnail detection, ranking penalties for weak visuals, capped filler backfill, and response metadata fields for modal rendering (mediaMode, embedUrl, previewBlockedReason).
- Expanded debug summaries with usable-thumbnail and embed counts, and added unit coverage for the new parsing/ranking helpers.
Why it changed:
- The latest production log showed Envato enrichment frequently failing, Artgrid enrichment collapsing on 403 plus HTML mismatch, and Gemini seeing too few usable visuals to do meaningful review.
How it was verified:
- go test ./...
What is still risky or incomplete:
- Frontend modal fallback behavior is not updated yet in this batch, so the new response metadata is not fully consumed until the UI patch lands.
- Envato source fetches may still fail on some pages if the provider changes challenge behavior again.
Date: 2026-03-16
What changed:
- Added in-process query translation / expansion cache inside GeminiService so repeated identical searches can reuse the same English query and variant list without re-calling Gemini or Google Translate.
- Added in-process response caching for repeated SearXNG requests and for source fetches used during Envato / Artgrid enrichment.
- Added in-process Gemini visual cache for fetched thumbnails and extracted preview frames so repeated candidate evaluation no longer re-downloads the same asset or reruns ffmpeg every time.
- Tightened backend tests to cover the new cache helpers.
Why it changed:
- Even after reducing query fan-out, repeated search passes and Gemini reevaluation were still paying duplicate translation, network, and media-processing cost inside the same running backend.
How it was verified:
- go test ./...
- bash scripts/selftest.sh
What is still risky or incomplete:
- These caches are process-local and TTL-based, so they do not survive container restarts and can still grow until expiry under heavy query diversity.
- Cache invalidation is intentionally simple; if upstream preview assets change rapidly, stale data can be served for a few minutes.
Date: 2026-03-16
What changed:
- Result modal layout was rebuilt to match a top 16:9 embedded viewer with bottom-left full AI note and bottom-right action panel.
- Google Video results now load YouTube embed URLs in the modal viewer and keep the white Direct Download action in the lower-right panel.
- When Gemini evaluation comes back mostly negative or too weak, the backend now runs one supplemental search pass with broader intent variants and reevaluates the merged pool.
- Failed Gemini batch evaluations now retry sequentially candidate-by-candidate with a short delay so more candidates can still be processed when batch/token evaluation is unstable.
Why it changed:
- The requested modal information hierarchy was different from the previous implementation.
- The user wanted negative Gemini feedback to trigger more exploration instead of stopping at the first pool.
- Batch-level Gemini failures were causing too many results to skip evaluation entirely.
How it was verified:
- code-path inspection against the updated modal wiring and search flow
What is still risky or incomplete:
- Non-YouTube third-party pages can still refuse iframe embedding via CSP or X-Frame-Options.
- Sequential Gemini retries improve coverage but also increase latency when the model is degraded.
Date: 2026-03-16
What changed:
- Bumped frontend asset version and added result-modal initialization guards to avoid click failures when browser cache serves mismatched HTML/JS.
Why it changed:
- The result modal stopped opening after the modal markup refactor, which is consistent with stale cached frontend assets or partially initialized modal DOM.
How it was verified:
- static code inspection of modal DOM/JS bindings
What is still risky or incomplete:
- Browser cache behavior itself was not fully reproduced here, so a hard refresh may still be needed in an already-open client session.
Date: 2026-03-16
What changed:
- Envato preview extraction now also inspects INITIAL_HYDRATION_DATA when direct page meta / JSON-LD preview URLs are missing.
- Search result cards and result modal now surface AI note before source summary text.
- Google Video direct download action moved into the AI note area and now seeds Zone C input before opening the shared preview/download modal.
- Result modal no longer depends on third-party iframe embedding and instead shows internal preview media plus external-open fallback.
- Search flow now shuffles collector query order and lightly randomizes the top merged results to reduce identical repeated outputs.
Why it changed:
- Some Envato items still missed preview URLs.
- Third-party iframe embedding was failing for blocked sites and creating a poor modal experience.
- The user wanted AI note to be the primary explanatory text and Google Video download action to be more obvious.
- Repeated searches returning the same ordering made the discovery experience feel too static.
How it was verified:
- go test ./...
- python3 -m py_compile worker/downloader.py scripts/mock_searxng.py
- bash scripts/selftest.sh
- added unit coverage for Envato hydration preview extraction
What is still risky or incomplete:
- Envato hydration structure could change again, so this fallback is still heuristic.
- Full browser-level validation of the revised result modal and button placement was not fully reproducible in this environment.
- Search randomness currently changes ordering and query traversal, but does not guarantee materially different source pools if upstream SearXNG returns a narrow candidate set.

Current Environment Variables

APP_ROOT
APP_ADDR
SQLITE_PATH
DOWNLOADS_DIR
FRONTEND_DIR
WORKER_SCRIPT
SEARXNG_BASE_URL
SEARXNG_GOOGLE_VIDEO_ENGINE
SEARXNG_WEB_ENGINE
GEMINI_API_KEY

Local Development Environment Notes

This machine started without go, pip, ffmpeg, yt-dlp, or node.
To unblock local development without root:
- installed Go 1.24.0 under ~/.local/tooling/go
- installed pip in user site with get-pip.py --user --break-system-packages
- installed imageio-ffmpeg in user site and linked its bundled ffmpeg into ~/.local/bin/ffmpeg
- installed yt-dlp standalone binary into ~/.local/bin/yt-dlp
- linked go and gofmt into ~/.local/bin
- appended ~/.local/bin and ~/.local/tooling/go/bin to PATH in ~/.bashrc
node is still not installed on this machine.
This is acceptable for the current repo because there is still no Node-based frontend build or lint workflow in-tree.
If future frontend work adds a Node toolchain, document the exact version and setup steps here before pushing.

Local Self-Test Workflow

Primary command:
- bash scripts/selftest.sh
What it currently verifies:
- Go formatting for touched backend files
- Python syntax for worker + mock SearXNG
- go test ./...
- backend binary build
- local app boot with temp SQLite/download dirs
- /healthz
- /api/search using local mock SearXNG
- /api/upload
Notes:
- search step now retries to reduce startup timing flakiness
- this is a smoke test, not a browser-level verification suite

Verified Locally In This Environment

user-local Go toolchain available on PATH
user-local ffmpeg available on PATH
user-local yt-dlp available on PATH
go build -o /tmp/ai-media-hub ./backend
go test ./...
Python syntax check for worker + self-test helper
local app boot / /healthz through scripts/selftest.sh
local /api/search against mock SearXNG through scripts/selftest.sh
local /api/upload through scripts/selftest.sh
full browser-level validation was not fully reproducible in this environment
Node-based frontend lint/build step is still unavailable in this environment

Recent Change Log

Date: 2026-03-16
What changed:
- Added /api/preview/stream so remote preview assets are fetched through the backend instead of relying on the browser to load Envato / Artgrid media directly.
- MP4 previews are cached on disk under the downloads area, and HLS playlists are rewritten so segment fetches also flow through the same backend proxy route.
Why it changed:
- Direct browser loading of remote preview URLs was still unstable and often failed due to upstream restrictions or missing headers.
How it was verified:
- code-path inspection of preview proxy and playlist rewrite flow
What is still risky or incomplete:
- HLS caching is not yet persisted segment-by-segment; current implementation rewrites playlists and proxies segment requests live.
Date: 2026-03-16
What changed:
- Relaxed final recommendation merge so Gemini-reviewed non-negative items can still appear, and only a small preview-capable ranked filler set is used when the result list is otherwise too thin.
Why it changed:
- The stricter recommended=true only merge made the visible result set collapse too aggressively in real searches.
How it was verified:
- go test ./...
- bash scripts/selftest.sh
What is still risky or incomplete:
- Filler results can still appear with the fallback reason when Gemini-reviewed positives are scarce, though the amount is now intentionally capped.
Date: 2026-03-16
What changed:
- Google Video embed URL now uses youtube-nocookie with explicit origin to reduce player load failures.
- Gemini Vision prompt now forces a Yes or No verdict per candidate and only Yes candidates are merged into the final result set.
- Candidate visual fetch now prefers Envato / Artgrid preview video frames before thumbnails and sends a Referer header for thumbnail fetches.
- Envato / Artgrid enrichment now retries preview extraction once after a short delay when the first fetch still lacks a usable preview URL.
Why it changed:
- The user reported YouTube player error 153, too many fallback-style results, and source pages that appear to need additional loading time before preview URLs become visible.
How it was verified:
- log review from ai-media-hub-2026-03-16T05-05-58-704Z.log
- go test ./...
What is still risky or incomplete:
- If a YouTube video has embedding disabled by the uploader, no embed URL variant will fully bypass that restriction.
- Delay-and-retry HTML fetching cannot execute JavaScript, so pages that only expose preview URLs after full client-side rendering may still need a real browser-based fetch path later.
Date: 2026-03-16
What changed:
- Reduced the heaviest search-stage caps slightly: fewer query variants per request, smaller per-source result caps, lower enrichment scope, and a bounded Gemini candidate set.
Why it changed:
- The widened search configuration was pushing the request past the reverse-proxy timeout and surfacing 504 Gateway Time-out.
How it was verified:
- go test ./...
- bash scripts/selftest.sh
What is still risky or incomplete:
- Search coverage is still broader than the original baseline, but there is now an explicit tradeoff between result volume and request latency.
Date: 2026-03-16
What changed:
- Increased collector result caps and widened source-specific search query templates for Envato, Artgrid, and Google Video.
- Strengthened Gemini query-expansion and vision prompts with a professional video-editor framing.
- Restored result modal media fallback so Google Video uses YouTube embed while Envato and Artgrid can show preview video or thumbnail instead of blocked iframe pages.
- Expanded generic preview URL parsing so HTML-embedded .mp4 and .m3u8 sources are accepted more broadly.
Why it changed:
- Search result volume was too low.
- The user wanted Gemini to reason more like a professional editor.
- Envato iframe pages were being refused, Google Video modal opening was broken, and preview extraction still missed known media URLs.
How it was verified:
- local code inspection against attached Envato / Artgrid HTML samples
- go test ./...
What is still risky or incomplete:
- The attached Artgrid HTML sample is a generic homepage shell, so preview extraction still depends on what the live clip page or downstream assets expose at runtime.
- Some providers can still refuse iframe rendering even when Artgrid pages currently appear to work.

Unraid / Docker / CI Notes

Dockerfile uses:
- Go build stage
- static ffmpeg image stage
- Python runtime stage
Heavy apt ffmpeg install path was removed earlier to reduce build time.
Gitea workflow builds and pushes:
- git.savethenurse.com/savethenurse/ai-media-hub:latest
- git.savethenurse.com/savethenurse/ai-media-hub:${{ github.sha }}

Recent Relevant Commits

9637b76 Improve query intent handling and preview playback
6d9391b Expand Artgrid query coverage to artlist canonical URLs
d8cc32e Fix Gemini candidate cap causing search 500s
e426261 Fix Artgrid collector matching and split ranker
5aebbef Refactor search into source-specific collectors
ae091c5 Improve source parsing from Envato and Artgrid HTML
06ea4f3 Restore Envato and Artgrid fallback search breadth
7dfb1ad Stabilize search pipeline and improve preview diagnostics
6f3149a Add local self-test flow and fix fallback regressions
f968458 Rewrite TODO as project handover

Git / Push Status

Current branch in ongoing work: main
Current state:
- latest work in this environment has been pushed successfully multiple times after the earlier remote unpacker issue
- the older push failure note is historical context only and should not be treated as the current repo state
- on 2026-03-17, local commits 6e9d20b and 0495f34 were created during fresh-machine environment validation and documentation cleanup
- push from this machine was initially blocked by missing HTTPS auth helper configuration, but succeeded after explicit credentials were provided by the user
Operational note:
- because the frontend is static and aggressively cached, browser hard refreshes are often required after UI / modal / preview changes

Highest-Value Next Steps

Reduce /api/search latency further without collapsing result count
Rebuild the reverted search-expansion work from the previous stable baseline, but only after measuring where candidate quality collapses between ranked pool and final merge
Validate the reopened 5ca7aef search-breadth direction against real proxy timeouts and visible result count before widening it any further
Continue hardening Gemini batch parsing so truncated JSON and ignorable low-visual candidate failures do not inflate user-facing warning counts
Build a repeatable repo-local bootstrap script or documented setup command set for non-root machines so fresh PC setup does not depend on shell history
Improve Envato / Artgrid preview acquisition reliability so Gemini Vision sees real frames more often
Browser-verify the new result modal at multiple viewport heights and confirm translated Source Summary readability on real long descriptions
Evaluate whether the new Gemini supplemental-query generation is reducing irrelevant results on a small fixed benchmark query set
Measure runtime cost of live Artgrid preview transcoding and decide whether bounded in-memory throttling or concurrency caps are needed
If Artgrid playback is still required for previewless clips, investigate a browser-rendered fetch path or clip-page asset bundle analysis because static HTML + yt-dlp probe were both insufficient on the supplied sample
Revisit Google Video UX:
- current YouTube embed was abandoned due error 153
- current in-app panel is more reliable but less rich than a true embedded watch page
Build collector-specific integration tests with recorded SearXNG samples
Separate source enrichment tests from live network behavior using local fixtures
Add a browser-level preview validation path, especially for hover video and preview proxy routing
If Artgrid hover preview is still required, obtain one real clip HAR / DevTools network export and derive a stable preview URL parser
Build a small fixed real-query benchmark set to evaluate search quality before further tuning
If frontend tooling becomes available, add lint/build checks

Short Handover Summary

The codebase runs.
Upload / direct-download features mostly exist and are broadly usable.
Search is functional but still the least stable subsystem by a wide margin.
Envato source fidelity is better than before, but Cloudflare / fetch failures still affect enrichment and preview acquisition.
Artgrid source fidelity is improved, but query coverage and preview extraction are still unreliable.
There is now a local self-test workflow.
Backend debug logging is now much more detailed and intended to support exported log-file analysis from the in-app Logs panel.

Current Reality Check

Search request flow is now heavily instrumented.
The frontend Logs panel can capture:
- API request start / completion
- SearXNG request / response counts
- collector query expansion
- enrichment start / finish
- Gemini translation / vision preparation / batch failures
- preview proxy fetch / cache events
The latest broad issue pattern observed in logs is:
- too many SearXNG calls for a single request can still push total latency too high
- Envato / Artgrid often fail to provide enough preview-capable candidates for Gemini Vision
- Google Video is frequently the easiest source to retrieve and therefore can dominate final results
- YouTube embed error 153 made the prior Google modal approach unreliable

Active Problems

504 Gateway Time-out
- Root cause: /api/search can still become too expensive when query expansion, source collectors, enrichment, and Gemini batch retries stack together.
- Current mitigation: request-level time budget and partial-result return path.
- Residual risk: fewer reviewed results can be returned when the budget is exhausted.
Too many Google Video-only result sets
- Root cause: Envato / Artgrid queries can still produce repeated rawCount: 0 responses from SearXNG.
- Current mitigation: looser unquoted query variants for both sources.
- Residual risk: upstream SearXNG quality still dominates discovery.
Gemini Vision partial or weak evaluation
- Root cause: many candidates still lack usable thumbnails / preview frames, so Gemini sees fewer real visuals than the raw result count suggests.
- Current mitigation: more verbose visual-fetch logging, preview-video-first strategy for Envato / Artgrid, and partial backfill from ranked candidates.
- Residual risk: if source media cannot be fetched, Gemini quality still degrades sharply.
Envato / Artgrid preview instability
- Root cause: source HTML can be incomplete, fetches can fail, and some previews may only appear after client-side rendering or protected media access paths.
- Current mitigation: JSON-LD/meta/hydration parsing, delayed retry, preview proxy route, MP4 cache, and HLS playlist rewriting.
- Residual risk: a real browser-rendered fetch path may still be needed later for some pages.
Google Video popup UX
- Root cause: YouTube embed error 153.
- Current mitigation: dedicated in-app Google panel instead of direct embed.
- Residual risk: this is reliable but not as rich as showing the live watch page.

Current Technical Notes

Preview proxy route:
- /api/preview/stream
- MP4 responses can be cached to disk
- HLS playlists are rewritten so segment fetches also flow through the backend
Frontend cache busting is done via /app.js?v=...
If behavior in the browser does not match the latest backend/frontend code, the first assumption should be stale frontend assets until proven otherwise

Recent Change Log

Date: 2026-03-17
What changed:
- Restored hard/ignorable Gemini batch error filtering so low-value thumbnail skips and no-visual candidate skips do not count as user-facing partial batch failures when useful recommendations were still recovered.
- Added unit coverage for the Gemini batch error filtering behavior.
Why it changed:
- The user reported the warning gemini vision partially failed on 3 of 4 batches.
- The provided log ai-media-hub-2026-03-17T06-06-04-447Z.log showed three different batch-failure classes mixed together:
  - real Gemini JSON extraction failures
  - low-value thumbnail skips
  - no-visual candidate skips
- Only the first class should strongly influence the user-facing partial-failure warning.
How it was verified:
- go test ./...
- bash scripts/selftest.sh
What is still risky or incomplete:
- This reduces noisy partial-failure warnings, but it does not eliminate genuine Gemini JSON truncation or provider-side preview problems.
- If the model starts returning different malformed JSON patterns, the parser and warning logic may still need further hardening.
Date: 2026-03-17
What changed:
- Reapplied the broader-search / modal-fitting codepath from 5ca7aef as requested by the user.
- Added a targeted 504 mitigation by making Gemini batch recovery stop at the request deadline instead of continuing sequential single-candidate retries indefinitely.
- Kept an explicit Gemini vision JSON output token cap to reduce the chance of truncated model responses during larger structured batches.
Why it changed:
- The user explicitly asked to return to the 5ca7aef direction and then fix the live 504 Gateway Time-out.
- The provided log ai-media-hub-2026-03-17T04-59-24-566Z.log showed backend search collection finishing in about 22s, then Gemini batch failure plus sequential recovery continuing until the reverse proxy returned 504.
- The direct cause was deadline-unaware Gemini recovery work, not initial search collection itself.
How it was verified:
- go test ./...
- bash scripts/selftest.sh
- attempted docker build -t ai-media-hub:test . but docker is not installed in this environment
What is still risky or incomplete:
- This reduces one concrete 504 path, but it does not guarantee that all reverse-proxy timeout cases are eliminated under worse upstream conditions.
- Container-image build verification still needs to happen on a machine that actually has Docker available.
Date: 2026-03-17
What changed:
- Raised search breadth again by widening per-source collector caps, increasing the number of base queries considered, increasing per-collector query budgets, and expanding the Gemini candidate review budget.
- Added another browser-resolution-aware modal fitting pass: the result popup now recalculates shell width, shell height, media max height, and compact-mode behavior from window.innerWidth / window.innerHeight when opened or when the viewport changes.
- Removed the fixed minimum heights on the lower modal panels so they can shrink with the viewport instead of forcing the popup past the visible browser area.
- Checked the supplied Artgrid clip HTML and a direct yt-dlp probe against https://artgrid.io/clip/355470/...; the HTML still exposes metadata and thumbnail signals only, and the direct probe returned Unsupported URL.
Why it changed:
- The user reported that the popup still overflowed the browser viewport and that the visible result set still felt too thin despite the earlier search-expansion work.
- The attached Artgrid HTML sample was intended to verify whether a stable preview URL could be extracted directly from static clip HTML.
How it was verified:
- go test ./...
- bash scripts/selftest.sh
- python3 -m py_compile worker/downloader.py scripts/mock_searxng.py
- python3 worker/downloader.py --mode probe --url 'https://artgrid.io/clip/355470/couple-woman-man-holding-hand'
What is still risky or incomplete:
- The new modal layout now adapts to actual browser viewport dimensions in code, but a live browser confirmation on the exact user display is still needed.
- Widening search breadth and Gemini review budget should improve visible count, but it can also increase latency and token cost under poor upstream conditions.
- The supplied Artgrid HTML sample alone was not enough to derive a directly playable preview URL, so previewless Artgrid clips may still fall back to thumbnail-only behavior.
Date: 2026-03-17
What changed:
- If the first search pass plus Gemini filtering still leaves too few visible results, the backend now performs an additional coverage-expansion search/evaluation pass before final fallback filling.
- Search result cards and the result modal now reuse the existing direct-download preview probe path for Google Video, so YouTube-backed results can surface actual playable preview streams instead of staying thumbnail-only.
- Artgrid-style .m3u8 previews now have a server-side ffmpeg transcode route so the frontend can request a directly playable MP4 stream when plain HLS playback is unreliable.
- Persistent preview-file caching to disk was removed from the preview proxy path so one-off preview traffic does not keep accumulating files under the downloads cache area.
- Modal spacing, media height, and text sizing were tightened again so the popup is more likely to fit without scrolling on shorter displays.
Why it changed:
- The latest user feedback said the search result count had become too low after stricter Gemini filtering, the popup still felt too large, Google Video already had a better preview path in direct-download mode, Artgrid previews still were not reliably playable, and the server should prefer not retaining disposable preview artifacts on disk.
How it was verified:
- go test ./...
- bash scripts/selftest.sh
- python3 -m py_compile worker/downloader.py scripts/mock_searxng.py
What is still risky or incomplete:
- The extra coverage-expansion pass can improve visible count but may increase latency when upstream SearXNG quality is poor.
- Live ffmpeg transcoding avoids preview-file accumulation but may become CPU-heavy under concurrent preview playback.
- Real browser validation is still needed for the exact viewport in the latest screenshot and for actual Artgrid preview playback behavior.
Date: 2026-03-17
What changed:
- Removed the visible AI Recommended badge from search cards.
- Extended lazy summary translation so result cards, not just the modal, now request Korean Source Summary text as they come into view.
- Reworked Gemini vision parsing and recommendation metadata so candidate assessments now distinguish positive, unclear, irrelevant, and inappropriate, with short search hints for weak/negative results.
- Removed the direct Preview evidence pending filler path from normal merge behavior and moved fallback filling to a later, more neutral stage.
- Upgraded supplemental search behavior so follow-up queries can be generated from Gemini feedback and provider/source mix instead of relying only on the old fixed fallback list.
- Tightened modal height handling again so the overlay and internal panels can scroll without pushing the popup past the viewport.
Why it changed:
- The user reported that the modal could still overflow the browser, that the AI Recommended chip was unnecessary noise, that card-level source summaries should also be translated, and that weak/negative Gemini evaluations should trigger smarter additional searching instead of surfacing low-confidence filler results.
How it was verified:
- go test ./...
- bash scripts/selftest.sh
- python3 -m py_compile worker/downloader.py scripts/mock_searxng.py
What is still risky or incomplete:
- The new supplemental-query generation depends on Gemini availability for the smartest path and falls back to deterministic query building when Gemini text generation is unavailable.
- Card-level lazy translation reduces request burst compared with translating everything during /api/search, but still adds client-side translation traffic during scrolling.
- Real browser validation is still needed to confirm the modal no longer exceeds the viewport on the exact user display conditions shown in the screenshot.
Date: 2026-03-17
What changed:
- Added POST /api/translate/summary so the result modal can translate Source Summary text to Korean on demand with in-process caching.
- Reworked result modal sizing so the popup stays within the viewport and the summary area scrolls internally instead of stretching the whole dialog off-screen.
- Replaced the false-positive near-deadline warning heuristic with explicit deadline metadata from search collection / enrichment / Gemini stages.
- Added an Artgrid API 403 cooldown guard so repeated clip enrichments stop re-hitting the same blocked JSON endpoint for a while and fall back to HTML parsing directly.
- Added backend tests for timeout-warning gating, summary translation caching, and Artgrid 403 skip behavior.
Why it changed:
- The provided browser log showed a successful ~55s search still surfacing search returned partial results to avoid gateway timeout, and the user reported that the result popup could overflow the viewport and that untranslated / very long source summaries made the modal feel too large.
How it was verified:
- go test ./...
- bash scripts/selftest.sh
- python3 -m py_compile worker/downloader.py scripts/mock_searxng.py
What is still risky or incomplete:
- Full browser-level confirmation of the resized modal and lazy summary translation still needs live UI validation.
- The Artgrid 403 guard reduces repeated waste, but it does not solve the underlying upstream access restriction.
Date: 2026-03-17
What changed:
- Re-audited the repository structure and TODO.md handover state on a fresh machine.
- Bootstrapped a user-local development toolchain because the machine initially lacked go, pip, ffmpeg, and yt-dlp.
- Revalidated the documented self-test flow and updated this handover document with environment notes, current verification status, and next-step priorities.
- Committed the documentation update locally as 6e9d20b (Document fresh-machine setup and status) and 0495f34 (Record local-only push failure), then completed the delayed push after credentials were supplied.
Why it changed:
- Ongoing work will happen from different PCs, so the project needed an explicit record of what this repo requires to become runnable when root-level package installation is unavailable.
How it was verified:
- go version
- ffmpeg -version
- yt-dlp --version
- go test ./...
- bash scripts/selftest.sh
What is still risky or incomplete:
- This machine still has no node, so any future frontend build/lint workflow will need an explicit setup step before it can become part of local verification.
- The environment bootstrap done here is user-local rather than repo-automated, so reproducing it on another PC is still partly manual until a bootstrap script or setup doc lands.
- HTTPS push on this machine currently depends on manually supplied credentials, so a safer long-term credential helper or SSH setup is still worth documenting later.
Date: 2026-03-16
What changed:
- When Gemini batch evaluation is only partially successful, the API now explicitly backfills the final list with preview-capable ranked candidates instead of leaving the visible result set too thin.
- Result modal behavior was switched so Envato and Artgrid open their actual item webpages inside the modal iframe again, while Google Video still uses the YouTube embed path.
- Frontend asset version was bumped again to force clients off stale preview-handling code.
Why it changed:
- The new debug log showed recommendedCount: 2 plus one failed Gemini batch, but the final visible list still collapsed to 6 items even though the raw pool was 25.
- The user wants in-modal login/download for Envato and Artgrid rather than preview-only media mode.
How it was verified:
- log review from ai-media-hub-2026-03-16T06-00-31-359Z.log
- code inspection of post-Gemini backfill and modal rendering path
What is still risky or incomplete:
- Artgrid candidate volume is still often low because SearXNG is returning many 0-result queries for current Artgrid query templates.
- If a browser still holds older frontend assets, a hard refresh may be required before the latest modal/preview behavior appears.
Date: 2026-03-16
What changed:
- Significantly increased backend debug logging volume and routed it into the existing WebSocket debug stream so the in-app Logs panel captures much deeper request, search, preview, Gemini, and ranking traces.
- Added request-start/request-done API middleware logs with request body and truncated response body snapshots.
- Added service-level debug hooks for SearXNG requests, collector activity, enrichment passes, Gemini translation/vision steps, ranking batch stats, and preview proxy/cache behavior.
Why it changed:
- The user wants to export detailed browser-visible log files and use them later for root-cause analysis instead of relying on sparse summaries.
How it was verified:
- code inspection of debug hook wiring from backend services to WebSocket debug broadcasts
What is still risky or incomplete:
- Very verbose logging can make the Logs panel noisier and larger, especially during long searches or preview playback.
Date: 2026-03-16
What changed:
- Removed per-batch Gemini fallback injection so empty Gemini sub-results no longer automatically turn into many "Gemini Vision 응답이 부족해..." items.
- Relaxed final merge to keep more non-negative Gemini-reviewed candidates and allow a larger capped preview-capable filler set when the visible result list is too small.
- Slightly raised Envato / Artgrid caps and enrichment scope again after the stricter merge caused visible result count to collapse too far.
- Bumped frontend asset version to 20260316e so clients pick up the newer preview-proxy behavior.
Why it changed:
- The user still saw too few results and too many fallback-labeled items despite the earlier filtering changes.
How it was verified:
- go test ./...
What is still risky or incomplete:
- If the browser is holding an older cached app.js, a hard refresh may still be needed before the proxied preview path is actually used on the client.
Date: 2026-03-16
What changed:
- Added a request-level time budget so search collection, enrichment, supplemental exploration, and Gemini evaluation stop early enough to return partial results before the reverse proxy reaches 504 Gateway Time-out.
- Reduced query variant count and per-source caps slightly again to cut the worst-case number of SearXNG calls in a single search request.
Why it changed:
- The new debug log showed /api/search taking about 90s, with the timeout happening during the Gemini stage after a very large number of search and enrichment steps.
How it was verified:
- go test ./...
- bash scripts/selftest.sh
What is still risky or incomplete:
- Partial-result responses are now preferred over hard 504s, so some searches may return fewer reviewed items when the time budget is exhausted.
Date: 2026-03-16
What changed:
- Loosened Envato and Artgrid query templates by adding unquoted variants so source discovery is less brittle when SearXNG has weak exact-phrase coverage.
- Replaced the Google Video modal embed with a dedicated in-app panel using thumbnail + metadata + action buttons, instead of relying on unstable YouTube embed playback.
- Bumped frontend asset version to 20260316g.
Why it changed:
- The latest log showed Envato and Artgrid repeatedly returning rawCount: 0, which is consistent with over-constrained exact-phrase queries.
- Google Video still hit YouTube player error 153, so the modal needed a non-embed fallback that remains usable.
How it was verified:
- log review from ai-media-hub-2026-03-16T06-23-45-335Z.log
- go test ./...
What is still risky or incomplete:
- Even with looser queries, SearXNG quality still controls whether Envato and Artgrid candidates are discoverable for a given term.

54 KiB Raw Blame History

AI Media Hub Handover

Working Rule

Current State At A Glance

Current Status Summary

Current Architecture

Search Flow: Current Implementation

Direct Downloader Flow: Current Implementation

Major Work Completed So Far

Current Features Implemented

Important Current Constraints / Known Problems

Current Risks Around Search Quality

Frontend Debug Logger

Recent Change Log

Current Environment Variables

Local Development Environment Notes

Local Self-Test Workflow

Verified Locally In This Environment

Recent Change Log

Unraid / Docker / CI Notes

Recent Relevant Commits

Git / Push Status

Highest-Value Next Steps

Short Handover Summary

Current Reality Check

Active Problems

Current Technical Notes

Recent Change Log

54 KiB

Raw Blame History