LLM Integration

ASUN was designed from the ground up for use in LLM (Large Language Model) workflows. This page explains why it works better than JSON for structured model output.

The Problem with JSON + LLMs

When you ask an LLM to return structured data as JSON, several failure modes are common:

1. Trailing commas — invalid JSON but easy to write
2. Mismatched brackets — especially in deeply nested structures
3. Forgotten quotes — "name": Alice instead of "name": "Alice"
4. Repeated field verbosity — the model spends tokens repeating key names thousands of times

Even with function calling / tool use, JSON-based structured output has high failure rates on complex schemas.

Why ASUN Works Better

1. Less to get wrong

ASUN requires fewer special characters. Most string values are unquoted:

[{name@str, role@str}]:
  (Alice, admin),
  (Bob,   viewer)

No quotes around Alice, admin, viewer. No commas inside value strings unless escaped with \,. Substantially fewer syntax rules to violate.

2. Schema as a prompt header

Include the schema once in the system prompt or as a header:

Respond with ASUN using this schema:
[{id@int, name@str, sentiment@str, score@float}]

Example@[{id@int, name@str, sentiment@str, score@float}]:
  (1, Product A, positive, 0.92),
  (2, Product B, neutral,  0.51)

Now analyze the following reviews:
...

The model only needs to output the data rows — the schema is already established. Hallucination of field names disappears.

3. Token budget

With 65% fewer tokens than equivalent JSON, you can fit 3× as many data rows in the same context window. This matters enormously for batch analysis, few-shot examples, and RAG context injection.

Rough token comparison for 100 records

Format	Tokens (approx)
JSON	~3 000
ASUN	~1 050
ASUN-BIN (not LLM-usable)	—

4. Streaming-friendly

ASUN's row-oriented format means each tuple (...) is a self-contained unit. Clients can begin parsing results as they stream in — you don't need to wait for the closing } of a JSON array.

Recommended Prompt Pattern

## Output format
Use ASUN with the schema below. Output ONLY the ASUN block, no explanation.

Schema:
{field1, field2@int, nested@{child}, items@[str], ...}

## Data
...your input data...

For structured extraction with validation, use a library wrapper that:

1. Sends the schema in the prompt
2. Parses the ASUN response
3. Retries once on parse error (rare with ASUN)

Example: Sentiment Analysis Pipeline

import openai
import asun

SCHEMA = "[{id@int, sentiment@str, score@float}]"

def analyze(reviews: list[str]) -> list[dict]:
    numbered = "\n".join(f"{i+1}. {r}" for i, r in enumerate(reviews))
    prompt = f"""Analyze sentiment for each review.
Respond ONLY with ASUN using this schema:
{SCHEMA}

Reviews:
{numbered}"""

    resp = openai.chat.completions.create(
        model="gpt-4o",
        messages=[{"role": "user", "content": prompt}],
    )
    return asun.decode(resp.choices[0].message.content)