Your AI librarian
Not an oracle. Not a search engine. A librarian — that knows where your stuff is and hands it to the model when asked.
pip install librarian-mcp
“Two words of a seven-word title. She walked to the shelf, handed me the book — and three adjacent authors I hadn’t known I needed. That’s not search. That’s a librarian.”
Cut your AI costs 50%+ — up to 95%.
Haiku ties Opus accuracy at 1/19th the cost, when the Librarian hands it the right pages. Same answers, smaller model, smaller bill.
See the numbers ↓The First and Second Industrial Revolutions centralized production — big factories, one-size-fits-all. The 2nd Second Industrial Revolution de-centralizes it: 3D printers and local tools let any town make its own things. Your AI librarian does the same for knowledge. Your own library. Your own tool. Run it yourself. Every town its own fab shop — and its own librarian.
The Eyewitness Benchmark
| Vendor | Model | Tier | HOT accuracy | COLD accuracy | Δ (HOT−COLD) | HOT cost/Q | COLD cost/Q | HOT $/correct | HOT p50 latency |
|---|---|---|---|---|---|---|---|---|---|
| Anthropic | claude-haiku-4-5 | cheap | 98.7% | 5.3% | +93.4 | $0.0066 | $0.0009 | $0.0067 | 3.49s |
| Anthropic | claude-opus-4-7 | premium | 98.7% | 6.7% | +92.0 | $0.1272 | $0.0222 | $0.1289 | 5.87s |
| Perplexity | sonar-pro | premium | 98.0% | 9.3% | +88.7 | $0.0144 | $0.0034 | $0.0147 | 2.74s |
| gemini-2.5-flash | cheap | 94.7% | 12.0% | +82.7 | $0.0007 | $0.0002 | $0.0007 | 1.54s | |
| gemini-2.5-pro | premium | 94.0% | 8.7% | +85.3 | $0.0061 | $0.0016 | $0.0065 | 5.41s | |
| OpenAI | gpt-4o | premium | 93.3% | 8.7% | +84.6 | $0.0106 | $0.0015 | $0.0114 | 7.38s |
| Perplexity | sonar | cheap | 92.0% | 7.3% | +84.7 | $0.0041 | $0.0003 | $0.0045 | 3.14s |
| OpenAI | gpt-4o-mini | cheap | 89.3% | 11.3% | +78.0 | $0.0006 | $0.0001 | $0.0007 | 1.69s |
| Check | κ | n | Interpretation |
|---|---|---|---|
| Haiku vs Opus spot-check | 0.883 | 120 | Almost perfect agreement |
| Haiku vs Gemini cross-grader | 0.850 | 56 | Almost perfect agreement |
How to reference this
| Axis | Value (Chapter 1) | Audience |
|---|---|---|
| Chapter name | The Librarian | Marketing, landing copy |
| R-number | R9 | HN, Reddit, developer framing |
| Romulator code | 9000 | Researcher / citation / internal |
| Paper number | #48 Eyewitness | Academic, press |
| Semver | v0.2.0 | PyPI mechanical |
Try it — Romulator 9000
Select an intent to load a canonical memory packet from the preload. No install required.