Configuration

Overview

The self-hosted sovseal service reads settings from local environment variables, YAML config files, or in-memory dictionary overrides. This reference covers all required and optional configurations.

Required Environment Variables

Create and fill these in your edge Deno function configuration or container .env file.

Variable	Required	Default	Purpose
`OPENAI_API_KEY`	Yes	-	API credentials for OpenAI LLM/Embedder operations.
`JWT_SECRET`	Yes	-	Signs session tokens. Must be a long, secure random sequence.
`ADMIN_API_KEY`	Optional	-	Legacy admin key for backward compatibility.
`AUTH_DISABLED`	Optional	`false`	Set to `true` to disable auth for local testing. Never use in production.
`DASHBOARD_URL`	Optional	`http://localhost:3000`	Allowed origin for CORS validation.

Configuration Methods

Method A: YAML Config (`config.yaml`)

Pass settings structure using a standard YAML file layout:

vector_store:
  provider: qdrant
  config:
    host: localhost
    port: 6333
    collection_name: agent_memory

llm:
  provider: azure_openai
  config:
    api_key: ${AZURE_OPENAI_KEY}
    deployment_name: gpt-4-mini

embedder:
  provider: ollama
  config:
    model: nomic-embed-text

reranker:
  provider: cohere
  config:
    api_key: ${COHERE_API_KEY}

Load it inside your application SDK initializer:

from mem0 import Memory
memory = Memory.from_config_file("config.yaml")

Method B: In-Memory Dictionary Config

If initializing within Python code directly, define config options using dictionary key-values:

from mem0 import Memory

config = {
    "vector_store": {
        "provider": "qdrant",
        "config": {"host": "localhost", "port": 6333},
    },
    "llm": {
        "provider": "openai",
        "config": {"model": "gpt-4-mini", "temperature": 0.1},
    },
    "embedder": {
        "provider": "vertexai",
        "config": {"model": "textembedding-gecko@003"},
    }
}

memory = Memory.from_config(config)

Tuning Component Settings

1. Vector Store Indexes

Enable fast search responses by explicitly indexing metadata filter keys:

vector_store:
  provider: qdrant
  config:
    indexed_fields: ["category", "priority", "status", "user_id"]

2. LLM Extraction Temperature

Keep extraction temperature low (≤0.2) to ensure that fact extraction outputs remain deterministic and structured.

3. Reranker Depth

When using a Cohere/managed reranker, cap the search results returned by setting:

reranker:
  config:
    top_k: 10

Sending more results adds significant transit latency with diminishing improvements in ranking accuracy.

On this page