Skip to content

Latest commit

 

History

History
88 lines (60 loc) · 1.89 KB

File metadata and controls

88 lines (60 loc) · 1.89 KB

Benchmark results

Recorded measurements on the indexed codehelper repository. Deterministic local runs — no cloud LLM tokens.

Snapshot: July 2026 · ~689 files · ~7k symbols · 2,830 call edges


Call-graph caller lookup

40-symbol sample against textual ground truth.

Metric Value
Informative symbols (have callers) 31
No-caller agreement (true negatives) 9
Mean precision 0.968
Mean recall 0.968
Mean F1 0.968
Latency p50 / p95 2.8 ms / 4.0 ms

Symbol search (query, lexical)

BM25 + trigram + call-graph centrality. No embeddings.

Metric Value
Recall@1 0.988
Recall@5 1.000
Recall@10 1.000
MRR 0.994
Query latency p50 2.7 ms

Natural-language retrieval

8-query core regression set: CI requires Recall@1 ≥ 0.70.

13-query full set covers feature, bugfix, refactor, security, and architecture-style queries.


Response size (TOON vs JSON)

40-item query-style payload:

Format Bytes ~Tokens (÷4)
JSON (indented) 9,090 ~2,272
TOON (default) 5,311 ~1,327
Savings 41.6% ~945 tokens / response

vs blind file reads

40 locate-understand tasks:

Metric codehelper (query/context) Read whole files
Median tokens 14 4,488
Median tool calls 1
Token reduction 99.7%

Agent workflow: orchestrate vs manual MCP vs no MCP

13 indexed projects · 5 task kinds each · July 2026

Metric Orchestrate Manual MCP No MCP
Quality (avg score) 0.968 0.915 0.188
Agent-facing tokens / case 519 7,191 2,933
Latency / case 760 ms 394 ms 64 ms

Orchestrate quality by task kind

Kind Avg score
feature 0.99
refactor 0.97
explain 0.95
bugfix 0.95
dead_code 0.93