Merge pull request #3 from ClaudioDrews/fix/embedding-provider-agnostic

fix: make embedding provider-agnostic via env vars (closes #1)

.env.example

@@ -46,6 +46,17 @@ REFLECTION_LOG_PATH=/home/your-user/.hermes/logs/reflection_trigger.log
 # MaaS env path (reflection_trigger.py loads env from this location)
 MAA_ENV_PATH=/home/your-user/.env
 
+# ── Embedding Backend ─────────────────────────────────
+
+# Default: OpenRouter with Qwen3-Embedding-8B.
+# Recommended model: multilingual (excellent for non-English content),
+# high-quality 4096d embeddings, fast inference, affordable pricing.
+# To use a local model (Ollama, vLLM, llama.cpp), set both vars:
+#   EMBEDDING_API_BASE=http://host.docker.internal:11434/v1
+#   EMBEDDING_MODEL=nomic-embed-text
+# EMBEDDING_API_BASE=https://openrouter.ai/api/v1
+# EMBEDDING_MODEL=qwen/qwen3-embedding-8b
+
 # ── Strongly Recommended ──────────────────────────────
 
 # LLM extraction token limit — 1024 is too small, causes fabric truncation

docker/worker/services/embedding.py

@@ -1,5 +1,6 @@
 """
-Embedding client via OpenRouter.
+Embedding client. Provider-agnostic — defaults to OpenRouter, configurable
+via EMBEDDING_API_BASE and EMBEDDING_MODEL for local Ollama/vLLM/llama.cpp.
 Mandatory dimension validation.
 """
 import os
@@ -11,13 +12,17 @@ logger = logging.getLogger("cognitive-worker.embedding")
 
 OPENROUTER_API_KEY = os.environ.get("OPENROUTER_API_KEY", "")
 EMBEDDING_DIMS = int(os.environ.get("EMBEDDING_DIMS", "4096"))
-EMBEDDING_MODEL = "qwen/qwen3-embedding-8b"
-API_BASE = "https://openrouter.ai/api/v1"
+EMBEDDING_API_BASE = os.environ.get(
+    "EMBEDDING_API_BASE", "https://openrouter.ai/api/v1"
+)
+EMBEDDING_MODEL = os.environ.get(
+    "EMBEDDING_MODEL", "qwen/qwen3-embedding-8b"
+)
 
 
 async def get_embedding(text: str) -> list[float]:
     """
-    Generates embedding via OpenRouter.
+    Generates embedding via the configured backend.
     Validates that the returned dimensions match EMBEDDING_DIMS.
     """
     if not OPENROUTER_API_KEY:
@@ -38,7 +43,7 @@ async def get_embedding(text: str) -> list[float]:
 
     async with httpx.AsyncClient(timeout=60) as client:
         resp = await client.post(
-            f"{API_BASE}/embeddings",
+            f"{EMBEDDING_API_BASE}/embeddings",
             headers=headers,
             json=payload,
         )

layers/05-qdrant.md

@@ -2,9 +2,20 @@
 
 > **Service:** Qdrant 1.17+ (Docker)
 > **Collection:** `knowledge_base` (4096d Cosine + BM25 sparse)
-> **Embedding:** Qwen3-Embedding-8B via OpenRouter
+> **Embedding:** Qwen3-Embedding-8B via OpenRouter (default; configurable)
 > **Endpoint:** `http://localhost:6333`
 
+## Why Qwen3-Embedding-8B
+
+The default embedding model is **Qwen3-Embedding-8B** for four reasons:
+
+1. **Multilingual** — strong performance across 50+ languages, including Portuguese, Spanish, and other non-English content common in real-world agent use
+2. **Quality** — 4096-dimensional embeddings with high semantic fidelity; consistently ranks near the top of the MTEB leaderboard
+3. **Speed** — fast inference at 8B parameters, suitable for hourly ingestion pipelines without bottlenecking the ARQ worker
+4. **Cost** — affordable via OpenRouter ($0.025/1M tokens at time of writing); a full wiki re-index costs cents, not dollars
+
+Users can switch to any OpenAI-compatible embedding API by setting `EMBEDDING_API_BASE` and `EMBEDDING_MODEL` in `.env`. If changing models, ensure `EMBEDDING_DIMS` matches both the new model's output and the Qdrant collection schema.
+
 ## What it stores
 
 All knowledge that benefits from semantic search — wiki pages, session transcripts, raw documents, technical references. Content is ingested via the continuous ingest pipeline (hourly) and the wiki agent (scheduled).
@@ -75,7 +86,7 @@ Redis queue (ARQ job)
     │
     ▼
 ARQ Worker (Docker)
-    │ embed via Qwen3-Embedding-8B (OpenRouter)
+    │ embed via configured backend (default: Qwen3-Embedding-8B)
     │ get_sparse_embedding() → BM25 (fastembed, local)
     ▼
 Qdrant upsert (with dedup check)

setup/install.md

@@ -7,7 +7,7 @@
 - Hermes Agent 0.14.0+ (tested on 0.15.2)
 - Python 3.11+
 - Docker 24.0+
-- OpenRouter API key (for embeddings and LLM extraction)
+- OpenRouter API key (or configured embedding backend — see [Layer 5: Qdrant](../layers/05-qdrant.md))
 - 16 GB RAM recommended (8 GB minimum)
 
 ## Installation
@@ -80,6 +80,10 @@ ICARUS_EXTRACTION_MAX_TOKENS=4096
 ICARUS_EXTRACTION_MODEL=deepseek/deepseek-v4-flash
 EMBEDDING_DIMS=4096
 
+# Optional — Embedding backend (defaults to OpenRouter)
+# EMBEDDING_API_BASE=https://openrouter.ai/api/v1
+# EMBEDDING_MODEL=qwen/qwen3-embedding-8b
+
 # Optional
 ICARUS_OBSIDIAN=1
 ICARUS_RESULT_MAX_CHARS=500