Your first chat

Once the stack is up, drive a chat through the API to verify everything's wired correctly. Two paths: OpenAI Python SDK or raw HTTP.

Via the OpenAI SDK

Bigwave is OpenAI-compatible. Point the SDK at your local URL:

import os
from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8000/v1",
    api_key=os.environ["BIGWAVE_ADMIN_TOKEN"],
)

# Non-streaming
resp = client.chat.completions.create(
    model="llama3.2:3b",
    messages=[
        {"role": "system", "content": "You are a precise assistant."},
        {"role": "user", "content": "What is the capital of Australia?"},
    ],
)
print(resp.choices[0].message.content)
print("completion_id:", resp.id)

Streaming works the same way:

stream = client.chat.completions.create(
    model="llama3.2:3b",
    messages=[{"role": "user", "content": "Tell me a joke."}],
    stream=True,
)
for chunk in stream:
    delta = chunk.choices[0].delta.content
    if delta:
        print(delta, end="", flush=True)

Via raw HTTP

curl -H "Authorization: Bearer $BIGWAVE_ADMIN_TOKEN" \
     -H "Content-Type: application/json" \
     http://localhost:8000/v1/chat/completions \
     -d '{
       "model": "llama3.2:3b",
       "messages": [{"role": "user", "content": "Hello!"}],
       "stream": false
     }'

The response wire shape matches OpenAI exactly. Every existing client library (langchain, llamaindex, instructor, openai-python, openai-node) works against Bigwave with one URL change.

Submit thumbs feedback

Bigwave extends the OpenAI surface with a feedback endpoint. Capture the completion_id from the response and submit 1 (thumbs-up) or -1 (thumbs-down):

import requests
import os

requests.post(
    f"http://localhost:8000/v1/chat/completions/{completion_id}/feedback",
    headers={"Authorization": f"Bearer {os.environ['BIGWAVE_ADMIN_TOKEN']}"},
    json={"rating": 1, "label": "great"},
)

The feedback row lands in the workspace's FeedbackStore. It becomes preference data the next time you build a DPO dataset.

RAG-augmented chat

Drag a folder of notes into the Knowledge tab in the Web UI, or ingest via the admin API. Then chat with bigwave_kb_id to ground the response:

resp = client.chat.completions.create(
    model="llama3.2:3b",
    messages=[{"role": "user", "content": "What did the notes say about X?"}],
    extra_body={"bigwave_kb_id": "notes", "bigwave_rag_top_k": 4},
)

The gateway retrieves the top-k chunks from the workspace's vector store, assembles them into a system message, and prepends. The response includes x-bigwave-rag-status and x-bigwave-rag-hits headers so you can verify retrieval ran.

Inspect what happened

Every chat lands a row in the inference log:

curl -H "Authorization: Bearer $BIGWAVE_ADMIN_TOKEN" \
     "http://localhost:8000/admin/inference?limit=10"

The Loop dashboard at http://localhost:8080 (Loop tab) shows the active adapter, the latest eval pass-rate (with Wilson 95% CI), the last-7-day feedback ratio, and recent training jobs — auto-refreshing every 15 seconds.

What's next

If you want to…	Read
Fine-tune on your own data	Train SFT in getting started
Understand how feedback becomes preference data	The loop concept
Wire Bigwave into Cursor / Continue	Coding-assistant use case
Deploy beyond a laptop	Deploy guide