Voyage AI — Embedding Provider

Voyage AI is the sole embedding provider for OpenClaw’s 44-DB memory fleet. Every memory chunks_vec column across all 44 SQLite DBs was generated by voyage-4 (70% weight in hybrid 70/30 vector+FTS5 search). All embedding calls route through portkey — never directly to Voyage — using virtual key voyage-embeddin-9dffd3 and Portkey config pc-opencl-22d508. Read this hub before adding agents, changing embedding dimensions, or switching the vector model.

Quick reference

Field	Value
Vendor	Voyage AI
URL	https://api.voyageai.com
KB doc	API
Auth method	Bearer (API key)
Auth credential	op://Aurora/Voyage/api_key
Cred-proxy port	n/a (pending B1-B6 ratification per nemoclaw-audit-2026-05-03)
Webhook port	n/a
Webhook handler	n/a
Webhook dedup table	n/a
Tunnel path	n/a
Outbound API base	https://api.voyageai.com/v1 (routed via Portkey)
Rate limits	2,000 RPM / 8M TPM (voyage-4)
Rate-limit action	429 → exponential backoff (3 retries via Portkey), Discord ops alert
Cost	voyage-4: $0.00006/1K tokens; 200M tokens/month free tier
Backup/recovery	Vendor-owned; embeddings are reproducible (re-embed if lost)
Discord alert channel	ops
Drift cadence	manual (on model change or memory DB schema change)
Status	production
Primary model	voyage-4 (1024-dim default)
Portkey config	pc-opencl-22d508 (6h semantic cache, 3× retry)
Virtual key	voyage-embeddin-9dffd3
Duplicate VK	voyage (duplicate — consolidation candidate per portkey hub)

⚠️ Mandatory insert: Fleet-wide memory coverage

Already powers 44 SQLite memory DBs (~5,568 chunks fleet-wide, verified 2026-05-01) at /home/opsadmin/.openclaw/memory/*.sqlite. Every chunks_vec column is a voyage-4 vector at 1024 dimensions. Hybrid search weight: 70% vector (Voyage) + 30% keyword (FTS5). Cross-link: supabase (chunks_vec storage context — note: memory DBs are SQLite, NOT Supabase; Supabase is the RDBMS for structured state). Phase E.2 vault-wikilink suggester will reuse this embedding stack for semantic vault link suggestions. Vendor independence note: Voyage is the only vendor in the embedding tier — zero fallback. If Voyage becomes unavailable, embedding-based memory retrieval degrades to FTS5-only (30% weight path). OSIL Phase 6 may add a secondary embedding provider; track via openclaw-self-improvement-layer-2026-05-03.

Components

• /home/opsadmin/.openclaw/portkey-proxy/proxy.js — routes embedding calls via Portkey config pc-opencl-22d508; does NOT directly call Voyage
• /home/opsadmin/.openclaw/workspace/integrations/llamaindex/portkey_voyage_embedding.py — PortkeyVoyageEmbedding class; batches at 128; sets input_type per call direction
• /home/opsadmin/.openclaw/memory/*.sqlite — 44 SQLite DBs containing chunks_vec (vec0 tables) holding 1024-dim voyage-4 vectors
• /home/opsadmin/.openclaw/tools/claude-memory.py — memory CLI (search/save/list/stats/delete); calls Voyage indirectly via save path
• [[sources/kb/voyage/API]] — rate limits, model specs, pricing, error codes (last_verified: 2026-03-04)
• [[wiki/system-map/osil/vm-osil-overview]] — OSIL memory layer future state map

How it’s used

• Trigger condition: Any memory_save MCP call or python3 claude-memory.py save invocation causes an embedding call via PortkeyVoyageEmbedding
• Workflow: Text chunk → PortkeyVoyageEmbedding._get_text_embedding() → Portkey pc-opencl-22d508 → Voyage voyage-4 → 1024-dim float vector → stored in chunks_vec (vec0 table)
• Query workflow: memory_search → _get_query_embedding() sets input_type="query" → cosine similarity against chunks_vec → top-k blended 70% vector + 30% FTS5 score
• Agents involved: All 44 agents with SQLite DBs; 7 have explicit Voyage headers in models.json: solara, aurora, atlas, main, betteracq, betterfiles, reri
• Failure mode: Voyage API 429 → Portkey retries (3×, exponential backoff) → if still failing, embedding write to pending_saves queue; FTS5-only search degrades gracefully
• Success criteria: Memory save returns chunk_id; chunks_vec rowcount increases by 1; semantic search returns cosine similarity > 0.7 for near-duplicate queries

Cross-links

Agents that touch this

• _summary — highest memory density; aurora.sqlite is primary search target
• _summary — acquisitions.sqlite anomaly: only 1 chunk (Phase 7.5 investigation pending)
• _summary — intel/research agent; heavy memory write cadence
• _summary — dispo.sqlite; buyer match context
• _summary — backfill agent seeding memory DBs fleet-wide

Skills that invoke this

• il-marketplace-pull — il-marketplace-pull v2; enrichment results saved to memory
• acq-analyze — conversation analysis outputs saved as memory chunks
• acquisitions-followup — follow-up context retrieved via memory_search

Plans that govern this

• openclaw-self-improvement-layer-2026-05-03 — OSIL Phase 6 may add secondary embedding provider; B13 IL AI replication uses embeddings
• openclaw-fragmentation-fix-2026-05-01 — Phase 7.5: acquisitions.sqlite 1-chunk anomaly investigation
• openclaw-vault-wikilink-optimization-2026-05-03 — Phase E.2: vault-wikilink suggester reuses Voyage embedding stack
• nemoclaw-audit-2026-05-03 — B1-B6 ratification will determine cred-proxy eligibility for this endpoint

Feedback rules

• feedback_no_plaintext_creds — op://Aurora/Voyage/api_key never inline
• feedback_verify_schema_before_designing — check vec0 schema before adding new SQLite DBs
• feedback_audit_before_architect — verify chunks_vec rowcount before assuming embeddings exist
• feedback_chokepoint_principle — G-CHOKEPOINT; embedding writes are NOT in tool_calls table (embedding tier separate from LLM tier)
• feedback_dual_write_required — pending_saves fallback queue for failed embedding writes

KB / source docs

• API — embeddings API, reranking, rate limits, pricing (last_verified: 2026-03-04)

System maps

• request-lifecycle — embedding call path shown in Lane 4 (memory write)
• vm-osil-overview — OSIL memory tier future state

Related: LLM provider tier cluster

Voyage is in the Memory/embeddings cluster (anchor: supabase). For LLM routing cluster context, see:

• portkey — anchor; routes all Voyage calls via pc-opencl-22d508; CHOKEPOINT-1 enforcement
• anthropic — LLM brain; Voyage is the embedding layer; both route through Portkey
• openrouter — LLM fallback tier; does NOT handle embeddings
• moonshot-kimi — LLM specialty tier; does NOT handle embeddings

Related: Memory/embeddings cluster

Anchor: supabase

Member	Role
supabase	Primary RDBMS; agent memory is SQLite-separate but Supabase hosts tool_calls, processed_webhook_events, deal tables
portkey	Routes all Voyage embedding calls; pc-opencl-22d508 config; virtual key voyage-embeddin-9dffd3
anthropic	LLM; feeds text to be embedded; SOUL.md agent context saved to memory post-conversation
1password	Vault Aurora holds op://Aurora/Voyage/api_key

Open issues / TODOs

• Duplicate Portkey virtual key voyage (vs voyage-embeddin-9dffd3) — consolidation candidate; tracked in portkey
• acquisitions.sqlite only 1 chunk — Phase 7.5 investigation per openclaw-fragmentation-fix-2026-05-01
• cred-proxy-eligible: pending — will resolve after B1-B6 ratification per nemoclaw-audit-2026-05-03
• Phase E.2 vault-wikilink semantic suggester will add significant embedding call volume — monitor rate limits

Recent activity

• 2026-05-03: hub created (W2-S7)
• 2026-05-01: 44 SQLite DBs and ~5,568 chunks fleet-count verified
• 2026-03-04: KB doc sources/kb/voyage/API.md last verified