Configuration

All configuration via environment variables. No config file required.

Server

Variable	Type	Default	Description
`DAKERA_ROOT_API_KEY`	string	—	Bootstrap root API key (super_admin scope). If unset, auth is disabled.
`DAKERA_HOST`	string	`0.0.0.0`	Bind address
`DAKERA_PORT`	u16	`3300`	REST API port
`DAKERA_GRPC_PORT`	u16	`50051`	gRPC port
`RUST_LOG`	string	`info`	Log level: trace/debug/info/warn/error

Storage

Variable	Default	Description
`DAKERA_STORAGE`	`memory`	Backend: `memory`, `filesystem`, `s3`, `minio`
`DAKERA_STORAGE_PATH`	`./data`	Data directory for filesystem storage
`DAKERA_S3_BUCKET`	—	S3 bucket name (3–63 chars)
`DAKERA_S3_REGION`	—	S3 region (e.g. `us-east-1`)
`DAKERA_S3_ENDPOINT`	—	S3-compatible endpoint URL (MinIO)
`AWS_ACCESS_KEY_ID`	—	AWS / MinIO access key
`AWS_SECRET_ACCESS_KEY`	—	AWS / MinIO secret key

Embeddings

Variable	Default	Description
`DAKERA_MODEL`	`bge-large`	ONNX embedding model loaded at server startup. Values: `bge-large` (1024d, server default), `gte-modernbert-base` (768d, 8192 tokens, MTEB 64.38 — v0.11.94+), `modernbert-embed-base` (768d, 8192 tokens, MRL-capable — v0.11.94+), `bge-small` (384d), `e5-small` (384d), `minilm` (384d). Changing this mid-deployment requires re-ingesting all stored vectors.
`DAKERA_MRL_DIMENSION`	—	Matryoshka dimension truncation (v0.11.94+). Set to `64`, `128`, `256`, `512`, or `768` to post-pool truncate and re-normalize ModernBERT embeddings. Unset = native dimension. Non-MRL models ignore this setting.
`DAKERA_INFERENCE_DEVICE`	`cpu`	`cpu` or `cuda`

Reranker

v0.11.95+ — Default cross-encoder upgraded from Xenova/bge-reranker-base (278M, English-centric) to onnx-community/bge-reranker-v2-m3-ONNX INT8 (568M, multilingual, 51.8 nDCG@10 BEIR). The model is downloaded from HuggingFace at startup and cached in $HF_HOME.

Variable	Default	Description
`DAKERA_RERANKER_MODEL`	`bge-reranker-base`	Cross-encoder reranker model (v0.11.95+). Default: `bge-reranker-base` (278M params, ~49 BEIR nDCG@10). Set to `bge-reranker-v2-m3` (568M params, 51.8 BEIR nDCG@10, +17.7% quality) to upgrade. Production is byte-identical when unset — default remains `bge-reranker-base`. Targets Cat3/temporal recall improvements.
`DAKERA_RERANKER_ONNX_FILE`	—	Path to a custom reranker ONNX file (v0.11.95+). Use to A/B between INT8 and FP16 quantizations of the reranker model. Unset = bundled model file.
`RERANKER_POOL_SIZE`	`4`	Number of pre-allocated ONNX cross-encoder sessions in the reranker pool. Increase for higher concurrency at the cost of memory.
`RERANKER_ONNX_BATCH_SIZE`	`16`	Mini-batch size for cross-encoder ONNX inference. Each mini-batch is padded to its own max seq_len, reducing memory waste vs full-chunk padding.
`RERANKER_MAX_CONCURRENT`	`6`	Maximum concurrent reranker calls. Requests beyond this limit receive an immediate `503 Overloaded` response rather than queuing indefinitely.

Memory & decay

Variable	Default	Description
`DAKERA_DECAY_ENABLED`	`true`	Enable memory importance decay over time
`DAKERA_DECAY_HALF_LIFE_SECS`	`604800`	Half-life for decay (default: 7 days)
`DAKERA_AUTOPILOT_ENABLED`	`false`	Enable AutoPilot memory lifecycle management

Retrieval tuning

Variable	Default	Description
`DAKERA_RRF_K`	`60`	Reciprocal Rank Fusion k-parameter
`DAKERA_ML_CLASSIFIER`	`false`	Enable ML-based query classifier
`DAKERA_TEMPORAL_INFERENCE`	`true`	Enable per-query temporal boost (requires ML_CLASSIFIER)
`DAKERA_INFERENCE_TEMPORAL_MULT_BETA`	`0.65`	Multiplicative temporal scaling factor
`DAKERA_ENTITY_VECTOR_SEARCH`	`true`	Enable entity-filtered second HNSW pass merged via RRF
`DAKERA_ENTITY_LINK_BOOST`	—	Entity-linked memory scoring boost (v0.11.95+). Set to `1` to enable. Adds a popularity-decayed additive boost to recall candidates sharing named entities with the query — improves Cat2 (multi-hop) recall. Production is byte-identical when unset.
`DAKERA_HNSW_CACHE_MAX`	`50`	Max HNSW indexes in LRU cache
`DAKERA_FUSION_STRATEGY`	—	Hybrid search fusion override (v0.11.94+). `rrf` or `minmax`. Unset = MinMax (default, byte-identical). Enables live A/B comparisons without code changes.
`DAKERA_QUERY_DECOMP`	`false`	Query decomposition (v0.11.93+). When enabled, complex queries are decomposed into sub-queries and fused via RRF, improving multi-step recall.

Recall scoring weights (CE-36, v0.11.91+)

The compound recall score formula is w_vec·relevance + w_imp·importance + w_rec·exp(−Δt/τ) + w_freq·frequency. Defaults are byte-identical to pre-CE-36 behaviour.

Variable	Default	Description
`DAKERA_SCORE_W_VEC`	adaptive	Vector relevance weight (0–1). Unset = adaptive per query type via `smart_scoring_weights()`.
`DAKERA_SCORE_W_IMP`	adaptive	Importance weight (0–1). Unset = adaptive per query type.
`DAKERA_SCORE_W_REC`	`0.13`	Recency (temporal proximity) weight (0–1).
`DAKERA_SCORE_RECENCY_TAU_HOURS`	`168`	Decay constant τ in hours for recency term. Range: 1–87600 (1 h … 10 yr). Default: 7 days.

Auth, rate limiting & observability

Variable	Default	Description
`RATE_LIMIT_RPS`	`100`	Requests per second
`RATE_LIMIT_BURST`	`50`	Burst capacity
`DAKERA_AUTH_ENABLED`	`true`	Enable API key authentication
`DAKERA_CORS_ORIGINS`	`*`	Allowed CORS origins (comma-separated)
`DAKERA_TRACING_ENABLED`	`false`	Enable OpenTelemetry tracing
`OTEL_EXPORTER_OTLP_ENDPOINT`	`http://localhost:4317`	OTLP collector endpoint
`DAKERA_AUDIT_LOG`	`false`	Enable audit logging

Observability & monitoring

Dakera exports OpenTelemetry traces and Prometheus metrics. Configure OTEL to send traces to Jaeger, Grafana Tempo, or any OTLP-compatible collector:

Variable	Default	Description
`DAKERA_TRACING_ENABLED`	`false`	Enable OpenTelemetry distributed tracing
`OTEL_EXPORTER_OTLP_ENDPOINT`	`http://localhost:4317`	OTLP gRPC collector endpoint (Jaeger, Tempo, etc.)
`OTEL_SERVICE_NAME`	`dakera`	Service name in traces
`OTEL_RESOURCE_ATTRIBUTES`	—	Additional resource attributes (e.g. `deployment.environment=prod`)
`DAKERA_METRICS_PORT`	`3300`	Port serving `/metrics` (Prometheus format)

Prometheus scrape config: target dakera:3300/metrics. 13 built-in metrics covering request latency, cache hit rates, vector index size, memory count, and replication lag.

Backup configuration

Scheduled backups create encrypted, compressed snapshots of all data. Backups include vector indexes, memories, knowledge graph, and configuration.

Variable	Default	Description
`DAKERA_BACKUP_ENABLED`	`false`	Enable scheduled automatic backups
`DAKERA_BACKUP_INTERVAL`	`86400`	Backup interval in seconds (default: 24 hours)
`DAKERA_BACKUP_PATH`	`./backups`	Directory for backup files (local or mounted volume)
`DAKERA_BACKUP_RETENTION`	`7`	Number of backup snapshots to retain
`DAKERA_BACKUP_ENCRYPT`	`true`	AES-256 encrypt backup files (uses `DAKERA_ROOT_API_KEY` as KDF input)

Backups use zstd compression (~5x reduction). Trigger manual backups via dk admin backup or POST /admin/backup.

Configuration done — time to deploy

Set your env vars, mount your volume, and deploy. The deployment guide covers Docker, Docker Compose, Kubernetes, and bare-metal — with production-ready config examples for each.

Deployment Guide → Security Hardening →

Stay sharp on agent memory

Benchmark results, SDK releases, and production patterns. Under 500 words per issue.

✓ You're in. First issue lands soon — watch for Dakera in your inbox.

Dakera Cloud — zero-ops deployment, guaranteed SLA, team dashboards. Early access open.Get early access →