$5 free for your first month. No card, start in seconds
tiyuvta inference
Menu
use case · Qwen

Give your tools a memory

Most tools forget you the moment the chat ends. A new hire asks the same question last quarter’s hire already answered. Support repeats itself. The model is fine. It just has nothing to stand on.

A first run

three commands

Export TIYUVTA_KEY from the console. The runnable pack is avifenesh/tiyuvta-use-cases.

shell · first runtiyuvta-use-cases
export TIYUVTA_KEY="YOUR_KEY"
pip install -r requirements.txt
python examples/agent_memory.py "Who runs the hosted endpoint, and how am I billed?"
Get a keyExample repo

This page is a working shape: store what matters, find the right note later, then let the model answer from that — not from a guess.

You do it with three Qwen models on one prepaid key. No second bill. No product-specific SDK to rip out later.

Get a key and run the example. Or read the story first.

Why Qwen, three times

one family · three jobs

One family, three jobs:

  1. Remember. Qwen3-Embedding-8B turns each note into a vector you keep on your side. You pay once to store it. After that, retrieval is yours.
  2. Pick the right note. Nearest is not always right. Qwen3-Reranker-8B reads the question against the shortlist and scores what actually helps.
  3. Answer. Qwen3.8 27B writes from those notes. If the notes are not enough, it should say so.

That is the whole product pitch. Embed, choose, answer. Same key as chat.

What this is not

honest limits

This is not a hosted memory product. Your notes stay in your database. We never see the file path. We only see the text you send.

It is also not a quality bake-off. We have run the shape against the live endpoint. We have not published a recall score for “your company’s Slack.”

What you do

write · ask · keep what is true

Write a note. A decision, a price, a name, a policy. Save the text and its vector.

Ask later. The question is turned into a vector the same way. You take a shortlist, the reranker ranks it, Qwen 27B answers from the survivors.

Keep only what is true. After a turn, you can extract facts worth keeping. Do not write every reply back. That is how a store learns its own mistakes.

The three calls, if you want them in your own app:

pythonembed · rerank · answer
# 1. store a note
client.embeddings.create(
    model="qwen/qwen3-embedding-8b",
    input=["Prepaid credit does not expire. No subscription, no minimum."],
    dimensions=1024,
)

# 2. find notes for a later question — prefix the question, not the stored note
q = client.embeddings.create(
    model="qwen/qwen3-embedding-8b",
    input=["Instruct: Retrieve stored notes that help answer this question\nQuery: How are we billed?"],
    dimensions=1024,
)

# 3. rerank the shortlist (at most 64 notes per call)
# POST https://api.tiyuvta.ai/v1/rerank
# model: qwen/qwen3-reranker-8b

# 4. answer from the notes that survived
client.chat.completions.create(
    model="qwen/qwen3.8-27b",
    messages=[{"role": "user", "content": notes + "\n\nQuestion: How are we billed?"}],
)

A 429 on embed or rerank means retry in a moment. Nothing was charged for that try. Details live on the embedding and rerank pages.

What it costs, roughly

one balance

You pay to write a note once (embed, input only). You pay a little to rerank a shortlist (input only). You pay Qwen 27B to answer. One balance. Prices are on Pricing.

Next