Build the first working local chat adapter

Now implement the shortest complete path for the local Ollama model router with fallback. Keep every intermediate value available for inspection; hiding it behind a heavy framework would make this lesson harder to reason about. The output should be deterministic for the fixture (or schema-deterministic when a model is involved). Only after this path works should you generalize.

Implement a thin adapter around the local HTTP API. Start with a tiny model tag for bring-up. Print model ID and content; keep the adapter swappable for tests with a fake transport.

The artifact's user-facing goal is specific: answer allowed tasks locally when hardware permits, otherwise follow an explicit fallback policy. Its accepted input is prompt, model tag, timeout, and policy flags for whether cloud fallback is allowed. Those statements are intentionally narrower than “build an AI system.” Narrow scope lets us inspect every input and expected result, and it prevents a toy result from being presented as a production claim. System shape for this chapter: Ollama manages model weights and a local HTTP inference process; a thin application adapter applies a system policy, timeouts, and structured logging; an evaluation harness compares quality, tokens per second, memory use, and privacy requirements before enabling cloud fallback. Keep model calls behind adapters, keep authorization and validation in deterministic code, and carry stable IDs and versions through every response. That separation lets you decide whether a bad result came from input handling, retrieval, inference, validation, or deployment. This page's job is the implementation step: now implement the shortest complete path for the local ollama model router with fallback. Setup baseline for the chapter (run once per machine, not secrets in git):

# Install Ollama from the official site for your OS, then:
ollama pull llama3.2:1b
python -m venv .venv && source .venv/bin/activate
pip install httpx

If hardware or a hosted provider differs, preserve the interface and expected behavior. Do not present provider syntax as universal—when a vendor adapter is unavoidable, keep it behind a thin boundary and test with a fake first. The deliverable is not “it ran once”; it is a reproducible artifact another developer can inspect, including expected output and one deliberate failure related to hanging local inference or silent privacy-violating fallback. Operationally, write down the owner of this stage, the command you ran, the observed output, and the next page's dependency on that output. If you cannot point to a file, fixture, metric, or config key, the stage is not done. Prefer small, reviewable increments: one contract, one path, one metric, one failure, one gate. When tradeoffs appear—latency versus quality, hit rate versus false hits, local privacy versus cloud quality—record both numbers instead of moving the threshold until the report looks green. The chapter ships only when evidence for golden task pass rate, tokens/sec, peak memory, and fallback rate within policy and a rehearsed recovery path exist beside local-first router with LOCAL_MODEL_ID and ALLOW_CLOUD_FALLBACK pins.

Save this as lesson.py and run python3 lesson.py. Prefer the standard library or the pinned packages from the setup block so the example stays reproducible.

def local_chat(prompt):
    # Offline stand-in for POST http://localhost:11434/api/chat
    return {"model":"llama3.2:1b","message":{"role":"assistant","content":"Local OK: "+prompt[:20]}}
print(local_chat("Summarize embeddings.")["message"]["content"])

Expected output: Local OK summary prefix. Exact floating-point formatting may vary slightly, but the asserted behavior must not. Read the output as evidence about this stage, not merely proof that the interpreter started.

Print intermediate IDs, scores, or statuses before the user-visible answer. Fluent wrong outputs usually start as a wrong intermediate. At the implementation stage, save the smallest failing fixture beside the expected result. Change one cause at a time and rerun the exact command printed above; that makes the repair reviewable and keeps this chapter's progressive artifact reproducible.

After the happy path works, freeze the fixture as a regression oracle before adding queues, caches, or fancy UX. For this implementation page, preserve the fixture and result as evidence for the next page. Label observations separately from conclusions: a passing assertion establishes the behavior it names, while broader usefulness requires the chapter's full evaluation set and stated operating limits. Primary metrics for the chapter remain golden task pass rate, tokens/sec, peak memory, and fallback rate within policy.