Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Chroma — AI-native embedding database. In-process, lightweight vector store with automatic embedding, metadata filtering, and full-text search. Simplest path from prototype to production RAG.
.claude/skills/mkurman-chromadb/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-13 | ✗→✓ | ▲ Improved | -39% | 0% |
| case-04 | ✗→✓ | ▲ Improved | -67% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -35% | 0% |
| case-12 | ✗→✓ | ▲ Improved | -53% | 0% |
| case-06 | ✓→✓ | ＝ Same ✓ | -17% | 0% |
Chroma is an AI-native embedding database optimized for RAG workflows. Lightweight, in-process, with automatic embedding via sentence-transformers, metadata filtering, and semantic search — no separate server required. Fastest path from prototype to production.
bashuv pip install chromadb
pythonimport chromadb client = chromadb.PersistentClient(path="./chroma_data") collection = client.create_collection(name="documents") # Add documents with metadata collection.add( documents=["Paris is the capital of France.", "Berlin is the capital of Germany."], metadatas=[{"country": "France"}, {"country": "Germany"}], ids=["doc1", "doc2"], ) # Query with filter results = collection.query( query_texts=["What is the capital of France?"], n_results=3, where={"country": "France"}, ) print(results["documents"][0])
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-13 | fail→pass | 8,437 | 3,995 | -53% | 1 | 1 | 0% | 1,273 | 776 | -39% | 0 | 0 | — |
case-01 | fail→fail | 8,500 | 9,277 | +9% | 1 | 1 | 0% | 1,560 | 1,576 | +1% | 0 | 0 | — |
case-02 | fail→fail | 12,857 | 7,955 | -38% | 1 | 1 | 0% | 2,430 | 1,796 | -26% | 0 | 0 | — |
case-03 | fail→fail | 11,275 | 9,459 | -16% | 1 | 1 | 0% | 2,070 | 2,050 | -1% | 0 | 0 | — |
case-04 | fail→pass | 7,381 | 1,619 | -78% | 1 | 1 | 0% | 1,271 | 418 | -67% | 0 | 0 | — |
case-05 | fail→fail | 9,385 | 4,444 | -53% | 1 | 1 | 0% | 1,780 | 1,021 | -43% | 0 | 0 | — |
case-06 | pass→pass | 7,172 | 4,892 | -32% | 1 | 1 | 0% | 1,132 | 940 | -17% | 0 | 0 | — |
case-07 | pass→pass | 11,221 | 4,257 | -62% | 1 | 1 | 0% | 1,931 | 925 | -52% | 0 | 0 | — |
case-08 | pass→pass | 11,668 | 4,800 | -59% | 1 | 1 | 0% | 2,135 | 1,040 | -51% | 0 | 0 | — |
case-09 | fail→pass | 17,810 | 8,477 | -52% | 1 | 1 | 0% | 3,116 | 2,022 | -35% | 0 | 0 | — |
case-10 | pass→pass | 2,842 | 2,415 | -15% | 1 | 1 | 0% | 279 | 497 | +78% | 0 | 0 | — |
case-11 | pass→pass | 6,573 | 2,544 | -61% | 1 | 1 | 0% | 935 | 640 | -32% | 0 | 0 | — |
case-12 | fail→pass | 6,793 | 2,504 | -63% | 1 | 1 | 0% | 1,175 | 553 | -53% | 0 | 0 | — |
case-14 | pass→pass | 9,752 | 2,531 | -74% | 1 | 1 | 0% | 1,339 | 606 | -55% | 0 | 0 | — |
case-15 | pass→pass | 8,507 | 1,737 | -80% | 1 | 1 | 0% | 1,351 | 489 | -64% | 0 | 0 | — |
case-16 | pass→pass | 4,728 | 3,035 | -36% | 1 | 1 | 0% | 759 | 618 | -19% | 0 | 0 | — |
case-17 | pass→pass | 5,870 | 1,837 | -69% | 1 | 1 | 0% | 1,047 | 468 | -55% | 0 | 0 | — |
case-18 | pass→pass | 1,894 | 1,593 | -16% | 1 | 1 | 0% | 227 | 449 | +98% | 0 | 0 | — |
case-19 | pass→pass | 15,930 | 8,872 | -44% | 1 | 1 | 0% | 2,857 | 1,942 | -32% | 0 | 0 | — |
case-20 | pass→pass | 12,212 | 7,910 | -35% | 1 | 1 | 0% | 2,365 | 1,858 | -21% | 0 | 0 | — |
case-21 | pass→pass | 15,080 | 11,709 | -22% | 1 | 1 | 0% | 2,756 | 2,512 | -9% | 0 | 0 | — |
case-22 | pass→pass | 10,718 | 7,560 | -29% | 1 | 1 | 0% | 1,991 | 1,440 | -28% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +18 percentage points is the difference between those two pass rates over the 22 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.