Text summarization

After transcription, Sbobinator can generate an LLM summary of the text in trascrizione.txt.

Providers

ID	Engine	Setup
deepseek	DeepSeek API	API key in Summary settings
openai	OpenAI	API key
gemini	Google Gemini	API key
claude	Anthropic Claude	API key
kimi	Moonshot Kimi	API key
local	Qwen2.5 GGUF (CPU)	python scripts/download_summary_llm.py + RAM ≥ 16 GB

Configuration: http://localhost:8501/settings/summary

API keys are saved in data/.secrets/summary_keys.json (or SBOBINATOR_*_API_KEY variables — see .env.example).

---

Length

Value	Behavior
auto	Proportional to the text
short	Short
normal	Balanced
detailed	Longer

---

Pipeline

1. NeMo transcription completed
2. unload_model() — frees ASR RAM
3. summarize() — chosen provider, map-reduce for long text
4. Saves riassunto.txt

If the summary fails: job completed, summary_error with the reason, transcript and SRT stay valid.

---

Long texts

Beyond the provider's context → map-reduce strategy (partial chunks + merge).

---

CLI

sbobina transcribe file.wav -s --summary-provider deepseek
sbobina transcribe file.wav -s --summary-provider local --summary-length detailed

---

UI

1. Sidebar → Engines and API keys (first-time setup)
2. Check Generate summary
3. Choose engine and length
4. Enqueue the file

---

Python dependencies

pip install -r requirements/local.txt

Includes: openai, anthropic, google-genai, truststore, llama-cpp-python, psutil, huggingface_hub.

On Windows, truststore is required for the cloud APIs (Python 3.13 SSL fix).

---

Deprecated

LexRank / IT5 (extractive / abstractive) — removed. See bug-fix/FIX-RIASSUNTO-LLM.md.