feat(middleware): EmbeddingProvider injection — #178 native-app readiness

[Dewinator]

Summary

• Wires an optional embed: (text) => Promise<number[]> callback through Absorber, Digester, and PrimeFetcher constructors.
• proxy.ts builds the callback once via createEmbeddingProvider() and passes it to all three. When PR feat(embeddings): LlamaCppEmbeddingProvider behind MYCELIUM_LLM_PROVIDER — #178 part 1 #187 lands, flipping MYCELIUM_LLM_PROVIDER=llama-cpp will route embedding, absorbing, digesting, and prime-fetching through llama-cpp uniformly — no further wiring needed.
• Falls back to the existing direct Ollama fetch when no callback is provided (back-compat for the existing test suite + any operator scripts).

Why

The three middleware classes each duplicated a private 11-line embed() that hit Ollama's /api/embed directly — bypassing the EmbeddingProvider factory abstraction. Without this PR, even after #187 lands, the middleware path still depends on Ollama. That blocks the Welle-1 native-app track (#176).

Side benefit

The shared OllamaEmbeddingProvider already does sampleForEmbedding (head + middle + tail truncation) for inputs over 6000 chars. The previous direct-fetch path silently truncated from the END — the embedding then only represented the first ~2048 tokens of long auto-absorbed/digested texts. Routing through the provider fixes this for free.

Out of scope

• Removing ollamaUrl / embeddingModel constructor options (they remain as fallback for tests + back-compat). A follow-up can require the callback once feat(embeddings): LlamaCppEmbeddingProvider behind MYCELIUM_LLM_PROVIDER — #178 part 1 #187 is merged and the provider abstraction is the sole path.
• Switching the user-facing /api/chat proxy to llama-cpp — that's feat(app): Spike 2 — node-llama-cpp embedding + chat bridge (drop Ollama dependency) #178 part 2 (chat bridge).

Pillar check

• Pillar 1 (no cloud dependency) — strengthened: middleware embedding paths now respect the in-process llama-cpp provider when the env says so.
• Pillar 6 (security) — unchanged: same single trust boundary (the provider) handles model selection, GGUF checksum verification (feat(security): GGUF SHA-256 verification — Pillar 6 follow-up to #187 #188/feat(security): MYCELIUM_LLAMA_REQUIRE_CHECKSUM=1 fail-closed — Pillar 6 follow-up to #188 #189) flows through it.

Test plan

• npm run build — clean
• npm test — 943 pass, 0 fail (was 942 + 1 skipped)
• 4 new tests in middleware-embed-override.test.ts:
  • absorber routes through the override
  • digester routes through the override
  • prime-fetcher accepts the option (constructor-level — runtime path is identical to absorber/digester)
  • absorber falls back to direct fetch when no override is provided
• Manual after feat(embeddings): LlamaCppEmbeddingProvider behind MYCELIUM_LLM_PROVIDER — #178 part 1 #187 merges: MYCELIUM_LLM_PROVIDER=llama-cpp → middleware proxy auto-absorb confirms llama-cpp is hit (no Ollama traffic)

🤖 Generated with Claude Code

[Dewinator]

Peer audit (autonomous tick, 2026-05-02)

Read PR top-to-bottom against mcp-server/src/services/embeddings.ts (current main) and PR #187's pending diff to createEmbeddingProvider. Cross-checked TS dimension defaults against SQL VECTOR(768) constraints in migrations 008/016/060/071/087.

Claims that hold up

• "sampleForEmbedding side benefit" is real. Verified by reading absorber.ts/digester.ts/prime-fetcher.ts private embed() on main — they call fetch(\"/api/embed\") with raw text, no truncation, no sampleForEmbedding. Routing through OllamaEmbeddingProvider.embed() does add head+middle+tail sampling for >6000-char inputs. Long auto-absorbed memories silently improve in semantic recall after this lands.
• "No further wiring needed when feat(embeddings): LlamaCppEmbeddingProvider behind MYCELIUM_LLM_PROVIDER — #178 part 1 #187 lands" is correct. Confirmed against PR feat(embeddings): LlamaCppEmbeddingProvider behind MYCELIUM_LLM_PROVIDER — #178 part 1 #187's diff to createEmbeddingProvider() — it dispatches on MYCELIUM_LLM_PROVIDER (ollama|llama-cpp|llamacpp) and throws on unknown values. After feat(middleware): EmbeddingProvider injection — #178 native-app readiness #190 + feat(embeddings): LlamaCppEmbeddingProvider behind MYCELIUM_LLM_PROVIDER — #178 part 1 #187 are both on main, flipping the env routes absorber/digester/prime-fetcher embed paths through llama-cpp uniformly, no edits to middleware needed.
• Dimension contract is consistent. TS default dimensions = 768 matches every VECTOR(768) column + RPC signature in the migrations. No silent INSERT-time crash hiding here.
• Backward-compat path preserved. When the embed callback is unset, the constructor's ollamaUrl + embeddingModel fields still drive a direct fetch — the absorber-fallback test (test N4: Auto-Absorb + Auto-Digest Lifecycle-Hooks #4) explicitly exercises this and confirms the existing test suite's port-1 failure-mode pattern still works.

Non-blocking follow-up notes

1. Constructor embeddingModel/ollamaUrl become test-only paths in proxy.ts after this PR. Whenever embed is provided (always in proxy.ts), the constructor's URL + model fields are dead code. Already named in your "Out of scope" — flagging here so the cleanup PR you mentioned can land cleanly once feat(embeddings): LlamaCppEmbeddingProvider behind MYCELIUM_LLM_PROVIDER — #178 part 1 #187 is on main and the abstraction is the sole runtime path.
2. Prime-fetcher constructor-only test is the right call for this PR's scope — flagging only that an end-to-end runtime test for PrimeFetcher.build() would need db.rpc stubbing (the rpc rejects at port 1 before embed() is reached). The absorber+digester runtime tests already prove the pattern; the constructor-shape test is sufficient peer-review evidence for me. A future test could swap the db: PostgrestClient for a stub that returns empty rows, then assert the override fires inside build().

Verdict

MERGEABLE as-is. No file conflicts with #185/#187/#188/#189/#192 (verified by the empirical merge train in #176 comment 4364403408 — all 7 PRs merged in sequence with zero conflicts and 998/999 tests green). Per the suggested merge order, #190 lands cleanly any time after #187 (the load-bearing dispatch).

🤖 Posted via the audit-loop-break heuristic (memory 7556ea69): zero-comment PR, smallest code surface in queue, cross-checked TS↔SQL once, capped notes at 2.