Deleting an episode does not delete its data. DELETE /api/podcasts/episodes/{id}
resolves episode.audio_file and unlinks that one MP3 — the enclosing
data/podcasts/episodes/<uuid>/ directory, holding clips/ (one MP3 per dialogue
segment), outline.json and transcript.json, is never touched. Failed and
/retry-replaced runs leak a directory as well. Nothing reaps any of it: this
deployment had 15 directories on disk against 4 episodes in the database.
That is an upstream gap (roughly a shutil.rmtree of audio_path.parent.parent in
that handler), not something configurable here. Deliberately not patched by
bind-mounting a modified router — that would fork app logic into a deployment
repo and rot silently against pull_policy: always. Local cleanup instead.
scripts/prune_podcast_data.py reconciles the episode list from the API against
the directories on disk and removes:
- orphaned directories (no corresponding episode), and
- clips/ of completed episodes, since the clips are intermediate output once
the final MP3 exists.
Dry-run by default; --yes applies, --keep-clips restricts it to orphans.
Two guards, both tested: it aborts when any job is running/pending (a running
job has no audio_file yet, so its working directory is indistinguishable from an
orphan) and it skips directories touched within the last 60 minutes (--min-age).
It also refuses to act if the API is unreachable, rather than guessing.
Deletion runs inside the container (docker compose exec … rm -rf): the container
writes as root, so the host user cannot remove those directories — a plain
host-side rmtree fails with EPERM after the first directory.
Verified on this deployment: freed 13 MB (10 orphaned dirs + clips of 3 episodes,
49 MB -> 36 MB); all four surviving episodes still stream byte-identical MP3s
from /audio afterwards.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Podcast generation failed twice at the audio stage, for two unrelated reasons
that look similar from the UI but need opposite fixes.
1) TTS/STT ran as nohup background processes and silently did not survive a
reboot. Podcasts then failed with "Failed to generate speech: All connection
attempts failed" (httpx.ConnectError) even though outline and transcript had
generated fine.
Both now run as systemd user units. The unit files are versioned in
services/systemd/ (using %h, not a hardcoded home) and installed by
scripts/start_services.sh, which also enables linger so they start on boot
without a login session. They pin GPU 2 by UUID, not by index: CUDA orders
devices "fastest first", so index 2 can resolve to the T600.
2) GPU 2 is shared by three processes (TTS, STT and the separate chatterbox-tts
MCP service on :9999), leaving ~12 GB of headroom. podcast_creator sends
TTS_BATCH_SIZE (default 5) clips concurrently, and since /audio/speech is a
sync FastAPI handler, they generated genuinely in parallel on one shared
model. Activation memory multiplied, the TTS process hit 16.7 GB and threw
torch.OutOfMemoryError, surfacing as "HTTP 500" from the endpoint.
tts_server.py now serializes generation behind a lock (GPU-bound work, so
parallelism buys no throughput — it only multiplies peak VRAM) and frees the
cache afterwards. TTS_BATCH_SIZE=1 keeps the client from queuing requests in
that lock and running into esperanto's 300s TTS timeout; ESPERANTO_TTS_TIMEOUT
is raised to 600s as headroom.
Verified: 5 concurrent /audio/speech requests all return 200 with GPU 2 peaking
at ~11.5 GB (was 16.7 GB for the TTS process alone), and the previously failed
episode now completes end to end — 38/38 batches, 10:56 min of audio, zero OOM.
Docs record both failure signatures side by side, since ConnectError (server
dead) and HTTP 500 (server alive, out of VRAM) have very different remedies.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two root causes, both fixed in the bind-mounted podcast prompt templates
(prompts/podcast/*.jinja):
1. Language: the episode profile's `language` field IS passed to both
podcast_creator templates as {{ language }}, but the stock templates never
reference it — so podcasts came out English (driven only by the English
stock briefings/speakers), then read aloud by the German-locked Chatterbox
TTS = "English with a German accent". Added a CRITICAL LANGUAGE REQUIREMENT
block keyed on {{ language }} to both templates; now `language: "de"`
actually forces German. Verified end-to-end: a full run produced a 44-line
all-German transcript + audio.
2. Invalid json output failures: qwen3.5:27b is a thinking model; on long
segments the <think> block ate the response-token budget and truncated the
JSON. Prepended /no_think to both templates (~3x faster, valid JSON).
Templates are bind-mounted read-only via docker-compose (directory mount, so
edits survive a container restart without inode-staleness). Bundled briefings
and speaker backstories were also translated to German to reduce drift.
Known limitation documented: feeding an entire book (~70k tokens) as podcast
content makes each of the 6 LLM calls take ~3.5 min and is unreliable; use a
shorter source or summary. Confirmed working with concise content.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The three bundled episode profiles (business_analysis, solo_expert,
tech_discussion) shipped with language=null, which podcast_creator treats as
"unspecified" and defaults to English for outline/transcript generation
regardless of source language. Set language="de" on all three.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
esperanto's OllamaLanguageModel silently caps num_ctx at 8192 unless the
credential sets it explicitly, no matter what OLLAMA_CONTEXT_LENGTH or the
model's real context window is. Any chat context beyond that got truncated,
and because qwen3.5:27b is a thinking model, the truncated prompt usually
produced an empty final answer (swallowed by clean_thinking_content) rather
than a visible error - looked like "chat gives no answer" in the UI. Set
num_ctx=98304 on the Ollama chat credential: the largest value that still
keeps the whole model on GPU 1 (131072 spills onto CPU, 5-6x slower).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
qwen3-embedding (repurposed 8B LLM, ~13GB loaded) didn't fit on GPU 1
alongside the resident 17GB chat model, so Ollama silently ran it 44-56%
on CPU. A single embed call took 39-54s, so any real document's chunk
batches blew past esperanto's 60s embedding timeout and got stuck retrying
forever - looked like "adding the source failed" but extraction was fine.
Also note: credential num_ctx doesn't work for embeddings (esperanto sends
it at the wrong JSON level, Ollama ignores it) - a small dedicated
embedding model is the real fix, not context-window tuning.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
male_thorsten/female_kerstin come from Piper's openly-licensed German voice
models, not scraped real-person audio, to avoid cloning someone's voice
without consent. Speaker profiles now use distinct voices per speaker instead
of everyone defaulting to the owner's own cloned voice.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
tts_server.py now resolves `voice` against services/voices/*.wav (gitignored,
personal recordings) and clones via chatterbox's audio_prompt_path, falling
back to "default" (symlinked to ~/chatterbox-tts-cli/my_voice_deutsch_60s.wav)
for unknown names, or to the old built-in-voice-by-language-code behavior if
`voice` is a recognized language. All podcast speaker profiles now point at
voice_id "default" until distinct per-speaker reference clips are added.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Wraps the locally installed chatterbox-tts and faster-whisper packages in thin
FastAPI servers implementing OpenAI's audio API shape, since neither Ollama nor
OpenRouter support speech. Pinned to GPU 2 to avoid contending with Ollama's
resident models on GPU 1. Requires UFW rules for 8901/8902 (same pattern as
the existing 11434 rule) since UFW defaults to deny-incoming.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
