You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Reference for the Dakera server environment variables used by the v0.12 compose files and manifests (Dakera v0.12.1) (v0.11: the release/0.11 branch). The v0.12 features, their variables and constraints are in docs/features-v0.12.md. Compose passes to the container only the variables its environment: lists: a name in .env that no file passes never reaches the server (the base files pass the ones below and in .env.example; the overlays pass their feature's). The authoritative list is the server's env-vars.md; a name the server does not read is reported at startup (dakera --check-config). Set these in docker/.env (Docker Compose) or in your Kubernetes Secret/ConfigMap.
Required (Production)
Variable
Description
DAKERA_ROOT_API_KEY
Root API key. Generate with openssl rand -hex 32. The default compose refuses to start without this.
MINIO_ROOT_USER
MinIO admin username. Change from minioadmin in production.
MINIO_ROOT_PASSWORD
MinIO admin password. Change from minioadmin in production.
Core
Variable
Default
Description
DAKERA_HOST
0.0.0.0
Bind address
DAKERA_PORT
3000
REST API port
DAKERA_GRPC_PORT
50051
gRPC API port
DAKERA_STORAGE
memory
Storage backend: memory (ephemeral), filesystem or s3 (persistent)
DAKERA_STORAGE_PATH
/data
Data root. WAL, knowledge graph, filesystem backend, hot tier ({root}/hot) and warm tier ({root}/cache/warm) derive from it; mount the data volume here
Hot-tier budget of tiered storage; no effect without DAKERA_TIERED_STORAGE. With a unit suffix (512MB, 1g) it is bytes; a bare number below 10 million is a vector count (10 million or more: bytes, with a warning). The compose files use 512MB, the ConfigMap 1GB
DAKERA_DISK_CACHE_DIR
—
Enables the L2 on-disk read cache (v0.11's DAKERA_L2_CACHE_PATH is no longer read)
DAKERA_CACHE_DIR
{root}/cache/warm
Warm-tier directory (tiered storage only)
Tiered Storage
Automatically moves data between hot (L1), warm (L2/RocksDB), and cold (L3/S3) tiers.
Variable
Default
Description
DAKERA_TIERED_STORAGE
false
Enable tiered storage (needs DAKERA_STORAGE=s3; the compose files and manifests turn it on). The hot tier is RocksDB by default in v0.12 (DAKERA_HOT_TIER=memory keeps the v0.11 tier; DAKERA_ROCKSDB_SYNC=false skips the per-write fsync)
DAKERA_HOT_TO_WARM_SECS
3600
Seconds before hot → warm tier transition
DAKERA_WARM_TO_COLD_SECS
86400
Seconds before warm → cold tier transition
DAKERA_AUTO_TIER
true
Automatic tier transitions
DAKERA_TIER_CHECK_INTERVAL_SECS
300
Interval between tier sweep checks
Cluster (HA Mode)
Variable
Default
Description
DAKERA_CLUSTER_MODE
false
Enable cluster mode
DAKERA_CLUSTER_ROLE
—
Node role: primary or replica
DAKERA_CLUSTER_SEEDS
—
Comma-separated seed nodes (host:port,host:port)
DAKERA_CLUSTER_SECRET
—
Required in cluster mode: shared secret of node-to-node routes and gossip, >= 16 characters, identical on every node. Keep it in a Secret / .env.ha, never in a ConfigMap
DAKERA_CLUSTER_NODE_ID
generated once
Stable node identifier (DAKERA_NODE_ID is its legacy alias)
DAKERA_GOSSIP_PORT
7946
Gossip protocol port
DAKERA_GOSSIP_BIND
0.0.0.0:7946
Gossip bind address
DAKERA_API_ADVERTISE
—
Advertised API URL for the node
DAKERA_REDIS_URL
—
Redis URL for the L1.5 cache, rate-limit counters and SSE pub/sub (v0.11's DAKERA_CACHE_REDIS_URL is no longer read). Not set by the compose files: one Redis shared by the nodes of a v0.12.0 cluster kept replicated writes from being applied
Models
Variable
Default
Description
DAKERA_MODEL
bge-large
Embedding model. bge-large and the reranker ship in the image; any other model downloads into the model cache (/app/models) on first use or via dakera models pull
DAKERA_ALLOW_MODEL_CHANGE
—
From 0.12.1 only needed where the background re-embed cannot run (DAKERA_TIERED=1, late-interaction scoring, the visual lane, a recorded model this build does not know): set 1 for ONE start to acknowledge that the store was embedded by another model, then re-embed with POST /admin/namespaces/migrate-dimensions and remove it. Otherwise a changed DAKERA_MODEL re-embeds the store in the background
DAKERA_TIERED
false (compose: 1)
The tiered embedding engine (not tiered storage): it always embeds with bge-large, ignores DAKERA_MODEL and refuses late interaction. Identical on every node sharing a store
A Hugging Face mirror or internal proxy of the Hub
HF_HUB_OFFLINE
—
1: never download; a missing file fails at once naming it (air-gapped)
HTTPS_PROXY / HTTP_PROXY / ALL_PROXY / NO_PROXY
—
Proxy for model downloads (http:// or socks4/4a/5/5h://; nothttps://). Include 127.0.0.1,localhost,minio in NO_PROXY
Operator directories holding the model files (offline); SHA-256-checked unless DAKERA_MODEL_PATH_SKIP_VERIFY=1
Multilingual (docker-compose.multilingual.yml)
Variable
Default
Description
DAKERA_FULLTEXT_LANGUAGE
en (overlay: multilingual)
BM25 analyzer: an ISO 639-1 code or English name, or zh/ja/ko/th/none/multilingual (no stemming). Stored with each index; from 0.12.1 an existing index built with another analyzer is re-analysed in the background at the next start
DAKERA_FULLTEXT_CJK_BIGRAMS
follows the language (off for en)
Character-bigram indexing of unsegmented scripts
DAKERA_QUERY_LANG
en (overlay: auto)
en, de, fr, es, it, pt, nl or auto: query routing patterns, temporal expressions, date extraction. Per request: lang
Attachments, speech to text, vision (docker-compose.multimodal.yml, docker-compose.vision.yml)
Variable
Default
Description
DAKERA_ATTACHMENTS
off
Attachment routes and speech to text; 501 FEATURE_DISABLED when off
DAKERA_ATTACHMENT_MAX_BYTES
26214400
Largest upload; over it 413
DAKERA_WHISPER_MODEL
whisper-base
Speech-to-text model (WAV only): whisper-base (multilingual, language auto-detected, default and recommended), whisper-tiny.en (English, lightest), whisper-base.en (English), whisper-tiny (multilingual, lightest) or whisper-small (multilingual, quality)
DAKERA_VISION
off
Image/page indexing and the visual recall lane (needs DAKERA_ATTACHMENTS; use a dedicated data root and bucket)
DAKERA_VISION_MODEL
colmodernvbert
The visual model
DAKERA_MEM_HIGH_WATER_FRACTION
0.85
Media jobs reserve memory against limit x this; waits up to 10 s, then 503 + Retry-After
DAKERA_MEM_BACKPRESSURE
on
0 turns the memory refusals off
DAKERA_HNSW_CACHE_TTI_SECS
3600
Idle time after which whisper, the visual model and GLiNER are unloaded (also the ANN / full-text index idle horizon)
Records, late interaction, RaBitQ, rerank
Variable
Default
Description
DAKERA_RECORDS
off
Record routes: one vector + named extra representations (docker-compose.records.yml)
Most candidates any recall sends to the cross-encoder; request field rerank_candidates is capped by it. Unset, from 0.12.3: the largest pool a top_k <= 10 recall builds; 0 = no cap (the earlier default)
DAKERA_RERANK_WARMUP
on
Load the reranker at boot (the exact words 0/false load it on first use)
DAKERA_RERANKER_MODEL
bge-reranker-v2-m3
bge-reranker-base (or base) selects the smaller reranker
Security
Variable
Default
Description
DAKERA_ENCRYPTION_KEY
—
AES-256-GCM at rest: 64 hex characters or a passphrase of at least 8; identical on every node. Rotation uses a keyring (POST /admin/encryption/rotate-key) and needs no change to this value
DAKERA_CLUSTER_SECRET
—
Cluster mode: >= 16 characters, identical on every node
DAKERA_GRPC_ENABLED
true
gRPC listener; v0.12 gRPC needs an API key in the call metadata when authentication is on
DAKERA_CONFIG_LENIENT
—
1: start on a fresh install although a value cannot be honoured or a DAKERA_* name is unknown (listed in /healthconfig_warnings)
Request Limits
Variable
Default
Description
DAKERA_MAX_BODY_SIZE
10485760 (10 MiB)
Max request body size in bytes. The compose files and the ConfigMap set 524288000 (500 MB)
DAKERA_REQUEST_TIMEOUT
300
Request timeout in seconds, the outer ceiling. docker-compose.yml sets 600, docker-compose.ha.yml and the ConfigMap 120
Server-wide rate limit (REST and gRPC). v0.11's RATE_LIMIT_RPS / RATE_LIMIT_BURST still work, deprecated
DAKERA_HTTP_HEADER_READ_TIMEOUT_SECS
30
Server v0.12.3. Seconds a client has to send a complete HTTP/1 request head on the REST port, counted from when the server waits for one (a new or an idle keep-alive connection, which is then closed); [1, 86400]. A proxy that keeps idle upstream connections longer (ingress-nginx 60 s, Traefik 90 s by default; docker-compose.ha.yml's Traefik is set to 25 s) must be set below it, or this raised above the proxy's. From 0.12.3 it also closes an idle HTTP/2 connection that sends no request
DAKERA_MAX_CONNECTIONS
10000
Server 0.12.3. Most open REST connections, [1, 1000000]; one over the cap is closed at once. Keep it below the file-descriptor limit
DAKERA_SHUTDOWN_DRAIN_SECS
8
Server 0.12.3. Longest REST drain at shutdown, [0, 3600] (SSE streams end at the signal). Plus up to 15 s for media jobs: keep the sum below stop_grace_period / terminationGracePeriodSeconds (30 s here)
DAKERA_ALLOWED_HOSTS
—
Server 0.12.3, read only with DAKERA_AUTH_ENABLED=false: Host names accepted besides localhost, *.localhost, IPs, single-label names and private-use names (*.internal, *.local, *.home.arpa: host.docker.internal, a Kubernetes service FQDN) (comma list, * = any); others get 421 HOST_NOT_ALLOWED (DNS-rebinding guard; /health*, /internal/* exempt). A stack without auth reached by a public dotted name must set it
Recall (server v0.12.3)
The compose files pass these only when they are set (an older image reports a name it does not know).
Variable
Default
Description
DAKERA_NEIGHBORHOOD_MODE
inline
Neighbourhood context of a recall that names no neighborhood_mode: inline (after the top_k results, marked retrieval: "neighborhood"), separate (in context_memories; memories holds at most top_k) or off
DAKERA_NEIGHBORHOOD_EVENT_TIME
off
Neighbourhood window on event time instead of ingest time
DAKERA_RECALL_RANK_DEPTH
10
Ranking depth floor [1, 1000]: the first k of a top_k=k recall equal the first k of a top_k=10 recall
DAKERA_RECALL_COLLAPSE_DERIVED
off
One result per source memory (sentence sub-memories collapsed; matched_text, matched_ids)
DAKERA_RECALL_EXCLUDE_SUMMARIES
off
Session summaries left out unless a request sets include_summaries: true
DAKERA_RECALL_SESSION_CAP
0
At most N results of one session in the head of list / aggregation recalls (0 = no cap)
DAKERA_SENTENCE_DECOMP
on
0: no sentence sub-memories derived on any write path
DAKERA_SUPERSEDES_EVENT_TIME
on
Supersedes edges ordered by the memories' event dates when both have one
Document import (server v0.12.3)
On by default. The compose files pass these only when set. Server guide: docs/documents.md in the server repository.
Largest source document (413). A proxy in front must accept bodies this large (ingress-nginx: proxy-body-size, 1 MB by default; k8s/ingress.yaml sets 500 MB)
DAKERA_DOCUMENT_MAX_TEXT_BYTES
20971520 (20 MiB)
Most extracted text (413)
DAKERA_DOCUMENT_MAX_CHUNKS
5000
Most chunks per document (413, before any write)
DAKERA_DOCUMENT_CHUNK_TOKENS
400
Default chunk budget, [32, 8192], clamped to the model's cap
DAKERA_DOCUMENT_CHUNK_OVERLAP_TOKENS
50
Default overlap, [0, 1024]
DAKERA_DOCUMENT_EXTRACT_TIMEOUT_SECS
60
Wall clock of one PDF / DOCX extraction (504 EXTRACTION_TIMEOUT in the job)
DAKERA_DOCUMENT_EXTRACT_MEMORY_MB
1024
Address-space limit of one dakera extract-text child
DAKERA_DOCUMENT_EXTRACT_CONCURRENCY
2
Extractions at once. The container's memory limit must hold the server plus CONCURRENCY x MEMORY_MB (2 GB by default): 12G in docker-compose.yml, 4Gi in k8s/ (lower the concurrency or raise the limit for PDF-heavy imports)
Throughput and safety (server v0.12.3)
Unset, each is sized by the server from the host (cores, container memory limit). The compose files pass them only
when set.
Variable
Default
Description
DAKERA_MAX_CONCURRENT_INFERENCES
half the cores, at least 4
Inference permits for document embeds, and separately for query embeds
DAKERA_ADMISSION_WAIT_SECS
20
How long a write or recall queues for an inference permit or recall slot before 429 + Retry-After (0 = refuse at once; at most half of DAKERA_REQUEST_TIMEOUT). Bulk loaders raise it together with DAKERA_REQUEST_TIMEOUT
DAKERA_INTERACTIVE_MIN_CORES
half the cores, at least 1
Cores ingest leaves to waiting searches
DAKERA_EMBED_PREPACK / DAKERA_RERANK_PREPACK
auto
ONNX weight prepacking for the embedder / cross-encoder (bit-identical results, faster, about 0.3 GB of private memory per core each). Auto: on when the server runs under a container memory limit and the copies fit in 20% of it (docker-compose.yml's 12G / 4 CPU limit turns both on); off without a container limit (docker-compose.ha.yml). 1 / 0 force it
DAKERA_RERANK_CACHE_ENTRIES
65536
Exact cross-encoder score cache (LRU, about 150 bytes an entry, no text kept); 0 = off
DAKERA_RERANK_PASSAGE_TOKENS
0
Most passage tokens per rerank pair (0 = as many as fit); a long passage is scored on its best window
DAKERA_HNSW_CACHE_BYTES
40% of the memory limit
Byte budget of the ANN index cache (DAKERA_HNSW_CACHE_MAX, default 10000, caps its entries)
DAKERA_MEM_TRIM_INTERVAL_SECS
60
Periodic return of freed memory to the OS after growth (0 = only under memory pressure)
DAKERA_LONG_CONTENT_WINDOWS
off
Overlapping token windows for text longer than the embedding model reads, so a fact past its cap is found by vector search
Let dedup / consolidation combine memories of different sessions
DAKERA_AUTOPILOT_TRASH_RETENTION_HOURS
168
Memories removed by dedup or consolidation stay restorable this long (GET /v1/memory/deduplicated); 0 = deleted outright
REEMBED_BATCH_SIZE / REEMBED_INTERVAL_SECS
50 / 300
Tiered re-embed (DAKERA_TIERED=1): memories per cycle and idle poll; a backlog drains back to back
The server raises its soft open-files limit to the hard one at startup (the compose files set nofile to
65536 / 1048576).
Docker Compose Port Overrides
The default and HA compose files expose different ports to allow co-deployment on the same host.
Default Profile (docker-compose.yml)
Variable
Default
Service
DAKERA_PORT
3000
Dakera REST API
DAKERA_GRPC_PORT
50051
Dakera gRPC
MINIO_API_PORT
9000
MinIO S3 API
MINIO_CONSOLE_PORT
9001
MinIO web console
PROMETHEUS_PORT
9090
Prometheus (monitoring profile)
GRAFANA_PORT
3003
Grafana (monitoring profile)
DASHBOARD_PORT
3002
Dashboard UI (dashboard profile)
DASHBOARD_IMAGE
ghcr.io/dakera-ai/dakera-dashboard:0.5.2
Dashboard image
DAKERA_SESSION_TTL_HOURS
12
Dashboard 0.4.0+: longest an operator sign-in lasts (idle sessions end after 2 h)
DAKERA_TRUSTED_PROXIES
—
Dashboard 0.5.2: addresses / CIDRs (comma or space separated) of the reverse proxies whose X-Forwarded-For names the client (nginx real_ip; the sign-in throttle counts that address)
HA Profile (docker-compose.ha.yml)
All HA ports use the HA_ prefix to avoid conflicts with the default profile.