mongodb-memory-mcp: Skill for Claude Code

The ingestion pipeline

validate  ->  chunk (pure)  ->  embed (contextual, grouped by document)  ->  persist

One implementation, in storeContent() in src/services/knowledge-service.ts. Every surface — MCP store_content, POST /api/content, the web UI — goes through it. There is no second ingestion path and there must never be one.

Each stage may assume exactly what the previous one guarantees, and nothing more. The failures that matter here are the silent ones: a violated invariant usually does not throw, it just degrades retrieval in a way no test notices.

---

Stage 1 — Validate

Where: parseInput(storeContentSchema, args, 'store_content') in the transport, before the service is called. In: untrusted arguments from an AI client. Out: StoreContentInput.

Guarantees the rest of the pipeline relies on:

• content is a non-empty string of at most MAX_CONTENT_CHARS (5,000,000) and contains at least one non-whitespace character.
• contentType is one of CONTENT_TYPES — the chunker switches on it and has no default branch worth relying on.
• tags are trimmed, lowercased and deduplicated. Filters compare against the normalised form; skipping this makes tags: ['MongoDB'] unfindable.
• metadata is JSON-serialisable, under 32KB, has no key starting with $ and no prototype-pollution key. It is persisted verbatim.
• chunkOverlapTokens < chunkSizeTokens when both are overridden.

What breaks if you skip it: the MCP SDK validates against z.object(storeContentShape) and cannot see the superRefine rules. Removing the handler's parseInput call silently disables every cross-field check while leaving the tool apparently working. Validation failures must be ValidationError — they log at warn with event input.validation_failed, which is what keeps a caller's bad argument distinguishable from an outage.

---

Stage 2 — Chunk

Where: chunkContent({ content, contentType, options }) in src/chunking/index.ts, called through runChunking() in the service. In: normalised content (normalizeContent has already stripped a BOM and converted CRLF to LF). Out: ChunkingResult { chunks, strategy, stats }.