0007: Unified Embedding API

proposals/0007-embedding-api.md

@@ -0,0 +1,72 @@
+---
+GEP: 0007
+Title: Unified Embedding API
+Discussion:
+Implementation:
+---
+
+# Unified Embedding API
+
+## Abstract
+
+This GEP discusses the unified embeddings api and its interactions with embedding model providers.
+
+## Motivation
+
+Many LLM applications have a semantic search feature within their architecture. To support these applications on Glide, a unified embedding api is necessary.
+This will allow applications to embed chat requests as part of a retrieval augmented generation (RAG) application flow.
+
+### Requirements
+
+- R1: Handles all provider specific logic
+- R2: Easily maintained
+- R3: API schemas must unify common request params (e.g. dimensions)
+- R4: API routes must unify common embedding endpoints/API
+
+## Design
+
+```yaml
+routes:
+    chat: /v1/language/{pool-id}/chat/ 
+    transcribers: /v1/speech/transcribers/{pool-id}/  
+    speech-synthesizer:  /v1/speech/synthesizers/{pool-id}/ 
+    multi-modal: /v1/multi/{pool-id}/multimodal/ 
+    embedding: /v1/embeddings/{pool-id}/embed/ 
+```
+
+#### User Request Schema for Embedding Route
+
+```yaml
+{
+ "message": "Where was it played?",
+ "dimensions": 1536
+}
+```
+
+#### Response Schema for Embedding Route
+```yaml
+{
+  "provider": "cohere",
+  "model": "embed-multilingual-v3.0",
+  "provider_response": {
+    "embedding": [
+        0.0023064255,
+        -0.009327292,
+        ....
+        -0.0028842222,
+      ],
+    "token_count": {
+      "prompt_tokens": 9,
+      "total_tokens": 9
+    }
+  }
+}
+```
+
+## Alternatives Considered
+
+[TBU, what other solutions were considered and why they were rejected]
+
+## Future Work
+
+- Could we abstract away the entire RAG architecture? A single endpoint that takes a chat message -> embeds -> text semantic search -> LLM -> response