8.3.1 Application Development Roadmap: API, Tools, State

LLM application development is not just an input box plus a model API. A real feature validates input, calls models, uses tools, keeps state, parses output, logs errors, and gives users a recoverable experience.

See the Application Loop First

The chapter upgrades one model call into a maintainable application loop: input, prompt/context, model, optional tool, validation, output, feedback.

Run a Tool Dispatch Check

Function Calling means the model proposes structured action arguments, but your application must validate and dispatch them.

model_output = {
    "tool": "search_docs",
    "arguments": {"query": "RAG citations"},
}

allowed_tools = {
    "search_docs": {"required": ["query"]},
    "create_ticket": {"required": ["title", "priority"]},
}

tool = model_output["tool"]
required = allowed_tools[tool]["required"]
validation_ok = all(name in model_output["arguments"] for name in required)

print("validation_ok:", validation_ok)
print("dispatch:", tool if validation_ok else "block")

Expected output:

validation_ok: True
dispatch: search_docs

Never execute tool calls directly from model text. Validate tool name, argument schema, permission, and failure path.

Learn in This Order

Step	Read	Practice Output
1	LLM API practice	Write a robust call wrapper with timeout and error handling
2	Framework basics	Split prompt, model, tool, memory, retrieval, and parser roles
3	Function Calling	Validate structured tool arguments before dispatch
4	Hugging Face ecosystem	Know when hosted, local, or browser-side models fit
5	Dialogue systems	Store session state, slots, memory, and user feedback
6	Document and template apps	Turn parsing, extraction, and generation into modules

Evidence to Keep

Keep this page’s proof of learning as a small evidence card:

Request

input, state, tools/context, and expected output contract

Validated Output

parser/schema or business-rule check result

Trace

model call, tool/function call, document parse, or dialogue state

Failure Check

invalid format, missing field, stale state, or wrong tool

Next Action

prompt, schema, state, API, or parsing improvement

Pass Check

You pass this chapter when you can build a small assistant loop that handles one API call, one optional tool call, one structured output, and one error path.

The exit mini project is a course Q&A and study-planning assistant that classifies the user request, optionally retrieves knowledge, returns structured suggestions, and logs feedback.

Check reasoning and explanation

1. A passing answer traces the full path from query to chunks, retrieval scores, cited evidence, answer, and fallback behavior.
2. The evidence should include retrieved passages, source metadata, a cited answer, and at least one empty-retrieval or wrong-retrieval case.
3. A good self-check explains whether a failure came from chunking, retrieval, ranking, prompt assembly, missing sources, or unsupported generation.