GenAI examples — complete flows

The generative-AI recipe documents each piece; here they combine into real flows, self-hosted end to end. Install the extras per example: uv add "tempest-fastapi-sdk[genai,genai-rag,genai-audio]".

1. Load the model only if it fits

Before downloading gigabytes, check; let the SDK pick the precision that fits:

import asyncio

from tempest_fastapi_sdk.genai import TextGenerator, recommend

MODEL = "Qwen/Qwen2.5-7B-Instruct"
rec = recommend(model_id=MODEL)          # tries bf16 -> int8 -> int4

if not rec.fits:
    raise RuntimeError(rec.reason + " " + (rec.suggestion or ""))

gen = TextGenerator(
    MODEL,
    device=rec.device,
    quantization=rec.dtype.value if rec.dtype.value in ("int8", "int4") else None,
    idle_unload_seconds=300,             # free VRAM between bursts
)


async def main() -> None:
    """Run this example."""
    answer = await gen.generate("Explain PIX in one sentence.")
    print(answer)


asyncio.run(main())

2. RAG over a PDF knowledge base

Index the PDFs once, answer grounded questions afterwards. In-memory here (dev); swap for PgVectorStore(db, dim=384) in production.

import asyncio

from tempest_fastapi_sdk.genai import Embedder, TextGenerator
from tempest_fastapi_sdk.genai.rag import InMemoryVectorStore, PdfReader, Retriever

rag = Retriever(
    Embedder("sentence-transformers/all-MiniLM-L6-v2", normalize=True),
    InMemoryVectorStore(),
)
gen = TextGenerator("Qwen/Qwen2.5-7B-Instruct", quantization="int4")


async def main() -> None:
    """Run this example."""
    # indexing (once, at startup or in a job)
    for pdf in ("manual.pdf", "faq.pdf", "policies.pdf"):
        await rag.index(PdfReader().chunks(f"/kb/{pdf}", max_chars=1500, overlap=150))

    # querying (cheap, per request)
    async def answer(question: str) -> str:
        context = await rag.retrieve(question, top_k=5)
        prompt = f"{context}\n\nAnswer only from the sources above.\n{question}"
        return await gen.generate(prompt, max_new_tokens=400)


asyncio.run(main())

3. Web-grounded answer (SearXNG)

No own corpus — search the web, extract, answer:

import httpx

from tempest_fastapi_sdk import HTTPClient
from tempest_fastapi_sdk.genai import TextGenerator
from tempest_fastapi_sdk.genai.rag import ContentExtractor, SearxngBackend, WebSearch

search = WebSearch(SearxngBackend("http://localhost:8080", http_client=HTTPClient()))
extractor = ContentExtractor(http_client=httpx.AsyncClient())
gen = TextGenerator("Qwen/Qwen2.5-7B-Instruct", quantization="int4")


async def answer_from_web(question: str) -> str:
    context = await search.retrieve(question, extractor=extractor, max_results=5)
    return await gen.generate(f"{context}\n\n{question}", max_new_tokens=400)

4. Fully local voice assistant

The full loop: audio → text → RAG/LLM → text → audio. Nothing leaves the box.

from tempest_fastapi_sdk.genai import TextGenerator
from tempest_fastapi_sdk.genai.audio import Language, SpeechToText, TextToSpeech

stt = SpeechToText("base")
gen = TextGenerator("Qwen/Qwen2.5-7B-Instruct", quantization="int4")
tts = TextToSpeech.for_language(Language.EN_US)


async def voice_turn(audio_wav_path: str) -> bytes:
    """Take speech, return speech."""
    heard = await stt.transcribe(audio_wav_path, language=Language.EN_US)
    reply = await gen.chat([
        {"role": "system", "content": "You are a concise assistant."},
        {"role": "user", "content": heard.text},
    ])
    return await tts.synthesize(reply)          # WAV bytes

Combine with RAG: swap gen.chat(...) for gen.generate(await rag.retrieve(heard.text)) to get a voice assistant that answers about your PDFs.

5. FastAPI transcription endpoint

from fastapi import APIRouter, UploadFile

from tempest_fastapi_sdk.genai.audio import Language, SpeechToText

router = APIRouter()
stt = SpeechToText("base")                      # load in the lifespan in prod


@router.post("/transcribe")
async def transcribe(file: UploadFile) -> dict[str, object]:
    result = await stt.transcribe(await file.read(), language=Language.EN_US)
    return {"text": result.text, "duration": result.duration}

Scale and economy

• Batch: wrap embedder._embed_many in a BatchScheduler to coalesce concurrent embeddings into one forward pass.
• Share models: ModelRegistry(max_models=2) reuses loaded models across call sites and unloads the LRU.
• Free VRAM: idle_unload_seconds + gen.unload_if_idle() in a @tq.interval(60) TaskQueue task.
• Embedding cache: Embedder(cache=...) skips the model for text already seen.

Full reference for each piece: Self-hosted generative AI.