Structured Data Encoder

Rust-Accelerated SDE for LLMs

Compress structured data into compact token sequences — 40–85% fewer tokens, no model retraining, no heavy dependencies.

---

sd-encoder is the standalone Structured Data Encoder from CLM, compiled in Rust and exposed as a Python extension. It encodes dicts, lists, and nested objects into compact token sequences that LLMs interpret with equal or better accuracy at a fraction of the token cost.

Install it on its own if you only need structured data encoding — no spaCy, no NLP stack, no unnecessary overhead.

Input	Typical Compression
Product catalogs	55–85%
Knowledge bases	40–75%
Business rules	50–80%
API responses	45–70%

---

Installation

pip install sd-encoder

No additional downloads required. Pre-built wheels are available for Linux (x86_64, aarch64), macOS (Intel, Apple Silicon), and Windows.

---

Quick Start

from sd_encoder import SDEncoderV2, SDCompressionConfig

config = SDCompressionConfig(preserve_structure=True, auto_detect=True)
encoder = SDEncoderV2(config)

catalog = [
    {"article_id": "KB-001", "title": "Reset Password", "content": "To reset your password...", "tags": ["security"]},
    {"article_id": "KB-002", "title": "Update Billing",  "content": "To update your billing...",  "tags": ["billing"]},
]

result = encoder.encode_validated(catalog)
print(result.compressed)
# {article_id,title,content,tags}[KB-001,Reset Password,To reset your password...,security][KB-002,Update Billing,To update your billing...,billing]

print(f"{result.compression_ratio():.1f}% reduction")
print(f"{result.n_tokens()} → {result.c_tokens()} tokens")

---

Configuration

SDCompressionConfig controls field selection, truncation, and structure preservation. All parameters are optional.

from sd_encoder import SDCompressionConfig, FieldImportance

config = SDCompressionConfig(
    # Field selection
    required_fields=["id", "title", "status"],      # always include these
    excluded_fields=["internal_notes", "raw_log"],  # always drop these
    drop_non_required_fields=False,                 # if True, emit only required_fields

    # Importance filtering
    auto_detect=True,                               # infer importance from field name/value
    importance_threshold=FieldImportance.MEDIUM,    # drop fields below this level
    field_importance={                              # explicit overrides
        "summary": FieldImportance.HIGH,
        "version": FieldImportance.LOW,
    },

    # Truncation
    max_truncation_length=300,                      # global string truncation
    max_truncation_mapping={                        # per-field truncation
        "description": 150,
        "content": 500,
    },

    # Structure
    preserve_structure=True,                        # encode nested objects inline
    default_fields_order=["id", "title", "status"], # pin ordering of known fields
)

Field Importance

FieldImportance controls the auto-detection threshold. Values are ordered — NEVER < LOW < MEDIUM < HIGH < CRITICAL.

from sd_encoder import FieldImportance

FieldImportance.LOW      # drop when filtering
FieldImportance.MEDIUM   # include by default
FieldImportance.HIGH     # always include unless explicitly excluded
FieldImportance.CRITICAL # never drop (ids, names, titles)

# Comparable
FieldImportance.HIGH >= FieldImportance.MEDIUM  # True
int(FieldImportance.HIGH)                       # 3

Auto-detection applies heuristics to field names and values when auto_detect=True:

Pattern	Detected importance
id, uuid, name, title	CRITICAL
status, priority, details	HIGH
description, type, channel	MEDIUM
source, version, metadata	LOW
_*, *_at, *_date	NEVER

---

Output

encode_validated runs compression then strips redundant whitespace and falls back to the original if the compressed output is larger.

result = encoder.encode_validated(data)

result.compressed        # str — the encoded token sequence
result.original          # original input, returned as Python dict/list
result.component         # "ds_compression"
result.n_tokens()        # estimated token count of original
result.c_tokens()        # estimated token count of compressed
result.compression_ratio()  # float — percentage reduction

# Validate manually if needed
result = encoder.encode(data)
result.validate_compression_ratio()  # fall back to original if compressed is larger
result.validate_compressed()         # strip redundant whitespace

Use encode directly when you want to inspect the output before deciding whether to validate.

---

Benchmarks

Run the Rust load benchmarks with:

make bench

The load benchmark measures:

• encode_payload_size: latency and rows/sec for catalog-style table payloads and nested ticket payloads at 10, 100, 1,000, and 5,000 rows.
• encode_parallel_load: aggregate throughput with 1, 2, 4, and 8 concurrent workers encoding 100-row payloads.

Criterion writes detailed reports under target/criterion/. Use the thrpt line for capacity estimates and the time interval for latency bounds on the tested machine.

For a compact table that answers "how long does compression take and what ratio do I get?", run:

make profile-load

This prints average latency, p95 latency, estimated original/compressed tokens, compression ratio, and compressed bytes for flat records, catalog tables, nested ticket bundles, and API-style responses.

---

Encoding Examples

Single object:

encoder.encode_validated({"id": "T-42", "title": "Login fails", "status": "open", "priority": "high"})
# {id,title,status,priority}[T-42,Login fails,open,high]

Nested object:

encoder.encode_validated({
    "user": {"id": "U-1", "name": "Ana"},
    "ticket": {"id": "T-42", "status": "open"}
})
# {user:{id,name},ticket:{id,status}}[U-1,Ana][T-42,open]

List of dicts (table encoding):

encoder.encode_validated([
    {"id": 1, "name": "Laptop",  "status": "active"},
    {"id": 2, "name": "Monitor", "status": "active"},
])
# {id,name,status}[1,Laptop,active][2,Monitor,active]

With field filtering:

config = SDCompressionConfig(
    required_fields=["id", "title"],
    drop_non_required_fields=True,
)
encoder = SDEncoderV2(config)
encoder.encode_validated({"id": 1, "title": "Test", "internal_log": "...", "raw": "..."})
# {id,title}[1,Test]

---

Relationship to CLM

sd-encoder is the engine behind the Structured Data encoder in CLM. If you need thread or system prompt encoding alongside structured data, install the full library with the sd_encoder extra:

pip install "clm-core[sd_encoder]"

sd-encoder is the right choice when:

• You only need structured data encoding
• You want to avoid the spaCy dependency
• You're deploying in a constrained environment
• You're integrating encoding into a Rust or polyglot pipeline

---

License

• MIT License — License © 2025-PRESENT ([MIT License](LMIT License)) (mailto:info@clm-lang.com))

---

Issues · Discussions · Contact