sjawhar/ceramic

Ceramic Search

Lexical (keyword-based) web search. Configured as a global MCP in ~/.config/opencode/opencode.json — invoke the ceramic_search tool directly (not via skill_mcp).

Quick Start

ceramic_search(query="OpenAI GPT-5 announcement 2025")

Rewrite natural-language questions into keyword queries first. Fire 2–4 variants in parallel for better recall.

Tool

| Tool | Purpose | | ---------------- | --------------------------------------- | | ceramic_search | Search the web (returns ranked results) |

When to Use

• Current events, recent news, time-sensitive facts
• Topics where built-in knowledge may be stale
• Specific entities: people, products, technical terms, legal text, dates

How It Works

Ceramic is lexical, not semantic. It matches exact keywords and phrases — no synonym expansion, no intent inference. Word order matters (cat house ≠ house cat).

Querying — Keywords, Not Sentences

Rewrite natural-language questions into keyword queries before searching.

| ❌ Don't | ✅ Do | | -------------------------------------------------- | ----------------------------------------------- | | What are the best ways to invest money? | beginner investing strategies stocks bonds | | Why is rent so high in California? | California rent increase housing shortage 2025 | | technology news | OpenAI GPT-5 announcement 2025 | | California laws | California tenant security deposit return law |

Rules:

• Use specific entities (names, products, dates, locations)
• Drop conversational filler (how do I, what are, why is)
• Include explicit synonyms when terminology varies (college university tuition)
• Avoid vague abstractions (technology trends, people's feelings)

Multi-Query Strategy (Recommended)

Search is cheap. Fire 2–4 keyword variants in parallel instead of crafting one perfect query.

Example — user asks about climate impact on agriculture:

ceramic_search("climate change agriculture impact")
ceramic_search("global warming crop yields")
ceramic_search("developing countries farming climate effects")

Aggregate and rank results across queries.

Limitations

• Synonyms not matched: BBQ ≠ barbecue, gym ≠ fitness center — include both if unsure
• Misspellings fail: exact spelling required
• Conversational queries fail: rewrite into keywords first
• English only (other languages coming)

When to Combine with Other Tools

Use a different tool (or LLM reasoning) when the task requires:

• Understanding intent rather than matching words
• Conceptual/semantic similarity
• Reasoning over the answer (Ceramic returns documents, not answers)