codeboost 0.2.0

External reasoning framework for LLMs. Cognitive drift + fieldmap knowledge injection. Qwen 7B +10% on HumanEval, zero training.

codeboost

External reasoning framework for LLMs. Zero training, zero fine-tuning. pip install and go.

Injects structured coding knowledge into any LLM prompt. Weak models gain the most — up to +19.5% on HumanEval.

Install

pip install codeboost

Quick Start

from codeboost import query

result = query("Find all pairs in sorted array that sum to target")
print(result.context)  # inject this into your LLM prompt

Inject into any LLM

from openai import OpenAI
from codeboost import get_context

client = OpenAI()
context = get_context("binary search on rotated array")

response = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[
        {"role": "system", "content": f"Use this knowledge:\n{context}"},
        {"role": "user", "content": "Implement binary search on rotated sorted array"},
    ],
)

Works with any LLM: OpenAI, Anthropic, Ollama, Groq, DeepSeek, Gemini, Mistral...

How It Works

Five-stage pipeline:

1. Gates — classify task into IPOD phase (Input/Process/Output/Data), detect constraints and edge cases
2. Drift — 6-field cognitive analysis (Logic, Spatial, Pattern, Transform, Compose, Metaphor)
3. Fieldmap — query 5,766-node knowledge graph with boosted keywords
4. Describer — generate natural-language approach description (small models understand text > labels)
5. Compact — optional 89% compression for tiny context windows

Task → Gates → Drift → Fieldmap → Describer → Context string → LLM

Benchmark Results (HumanEval, pass@1)

All models: Qwen 2.5 Coder family. Full 164 problems. Tests executed, not string-matched.

Model	Raw	+ Codeboost	Delta	Wins / Losses
1.5B	29.9%	49.4%	+19.5%	42W / 10L
3B	72.6%	76.8%	+4.3%	14W / 7L
7B	64.6%	73.2%	+8.5%	19W / 5L
14B	62.2%	73.8%	+11.6%	22W / 3L

7B + codeboost ≈ 14B raw — same accuracy, half the VRAM.

Hardness Analysis

Codeboost helps most on hard problems and rarely hurts easy ones.

Model	Hard (raw fail)	Boost rescues	Easy (raw pass)	Boost damages
7B	58 problems	19 solved (32.8%)	106 problems	5 broken (4.7%)
14B	62 problems	22 solved (35.5%)	102 problems	3 broken (2.9%)

On 14B: 12:1 rescue-to-damage ratio — codeboost rescues 35.5% of hard problems while only breaking 2.9% of easy ones.

Ablation (7B)

Condition	pass@1	Delta
Raw (no codeboost)	64.6%	—
Fieldmap only (no drift)	70.7%	+6.1%
Full (fieldmap + drift + describer)	73.2%	+8.5%
Compact mode	68.3%	+3.7%

Feedback Loop

Track what works, avoid what doesn't:

from codeboost import query, feedback

result = query("two sum problem")
# ... use result.context with your LLM ...

feedback("two sum problem", "positive")   # next query reuses good patterns
feedback("two sum problem", "negative")   # next query avoids bad keywords

CLI

pip install codeboost[cli]

codeboost query "sliding window maximum"
codeboost stats
codeboost serve --port 8000

API Server

pip install codeboost[server]
codeboost serve

# POST http://localhost:8000/v1/query
# GET  http://localhost:8000/v1/stats
# GET  http://localhost:8000/health

Who Is This For

• Developers running local models (Ollama, vLLM) who want better results without upgrading hardware
• Startups self-hosting inference — cut model size in half, keep accuracy
• Offline / air-gapped environments — no API calls, no internet needed
• Batch processing — 10K solutions with a small model + codeboost costs 10x less than API

License

MIT