AKADATA LIMITED
AKAMAN Context Reduction Benchmark
AKAMAN benchmark results showing how focused local documentation fragments reduce model context while preserving the installed source as authority.
AKAMAN
A small Linux C utility that retrieves the compact documentation fragment an AI agent needs from the host where it is running.
Context reduction: less context, the same source
The representative benchmark contains 13 queries. Across them, the documentation that would otherwise enter context totals an estimated 1,210,101 tokens and 4,840,396 bytes; AKAMAN returned an estimated 1,937 tokens and 7,723 bytes. That is 1,208,164 estimated tokens avoided — a 99.84% reduction.
| Query | Page | Source → result | Reduction |
|---|---|---|---|
bash | bash(1) | 104,475 → 112 | 99.9% |
gcc | gcc(1) | 366,562 → 130 | >99.9% |
gcc -fPIC | gcc(1) | 366,562 → 125 | >99.9% |
curl | curl(1) | 79,701 → 71 | 99.9% |
curl --retry | curl(1) | 79,701 → 318 | 99.6% |
rsync --delete | rsync(1) | 65,026 → 107 | 99.8% |
grep -r | grep(1) | 7,923 → 120 | 98.5% |
find -mtime | find(1) | 23,595 → 62 | 99.7% |
tar --strip-components | tar(1) | 10,959 → 29 | 99.7% |
ip link set | ip-link(8) | 22,366 → 219 | 99.0% |
ip route add | ip-route(8) | 11,340 → 216 | 98.1% |
nft masquerade | nft(8) | 60,306 → 234 | 99.6% |
open(2) | open(2) | 11,585 → 194 | 98.3% |
What the numbers mean
AKAMAN does not compress or rewrite the source. It extracts the matching fragment and returns it; the rest of the manual simply never enters model context. The token figures are a consistent comparative estimate, not a measurement of any specific model tokenizer.
- Source documentation
- 4,840,396 bytes
- AKAMAN returned
- 7,723 bytes
- Estimated source tokens
- 1,210,101
- Estimated returned tokens
- 1,937
- Estimated tokens avoided
- 1,208,164
- Context reduction
- 99.84%
Methodology
The estimate is deliberately simple and consistent across every query:
estimated tokens = ceil(bytes / 4)It is not any of the following, and should not be read as one:
- an exact model-tokenizer measurement,
- an exact GPU-utilisation measurement,
- a billing or invoice-level token-saving claim,
- proof of an exact production token saving.
Why the GCC example is striking
The full local gcc(1) manual is roughly 366,562 estimated tokens. A single targeted query such as gcc -fPIC returns about 125 estimated tokens — the option description alone. The source was not compressed or paraphrased; the irrelevant 366,000 tokens merely never reached the model. That is the whole idea: AKAMAN keeps the documentation authoritative and local, and lets the agent ask for exactly the line it needs.