# ------------------------------------------------------------------
# KnowledgeFS integration service (started by the default Dify Compose deployment).
# Copy to knowledge-fs.env and configure durable dependencies before enabling product traffic.
# The Dify product flag remains disabled by default and no public proxy route is added.
# Dify inner API connection values and integrated mode are injected by docker-compose.yaml.
# Physical object storage is owned by Dify's configured STORAGE_TYPE and credentials; do not copy
# provider endpoints, bucket names, access keys, or secret keys into this service.
# Optional feature flags and capacity tunables use safe runtime defaults and belong here only when
# an operator intentionally enables or overrides that feature.
# ------------------------------------------------------------------

# Required durable database. Do not point this at Dify's application database or reuse
# Dataset/Document tables. Apply KnowledgeFS migrations before starting product traffic.
DATABASE_URL=
KNOWLEDGE_DOCUMENT_COMPILATION_RUNTIME=on
# A quiescing step performs several bounded cleanup statements in one transaction. Keep this
# comfortably below the derived worker lease while allowing normal database round trips to finish.
DURABLE_DELETION_STEP_TIMEOUT_MS=30000

# Source reads, parsing, archive/media extraction, thumbnails, and artifact writes share this
# process-wide admission budget across every supported document format.
KNOWLEDGE_DOCUMENT_MATERIALIZATION_MAX_CONCURRENCY=2
# A canonical ParseArtifact stays live through outline, semantic, graph, and index consumers.
# These process-wide limits prevent the compilation claim batch from retaining many large parser
# responses at once. One artifact charged at the byte ceiling is allowed to run exclusively.
KNOWLEDGE_DOCUMENT_RETAINED_ARTIFACT_MAX_CONCURRENCY=4
KNOWLEDGE_DOCUMENT_RETAINED_ARTIFACT_MAX_BYTES=134217728

# PDF image elements are rasterized by the Poppler executable bundled in the API image. The
# limits are per document/asset operation. Set the mode to off as an emergency or low-resource
# kill switch; do not set a host-specific command path when using the published image.
KNOWLEDGE_PDF_RASTERIZER=poppler
KNOWLEDGE_PDF_RASTERIZER_DPI=144
KNOWLEDGE_PDF_RASTERIZER_THUMBNAIL_DPI=48
KNOWLEDGE_PDF_RASTERIZER_TIMEOUT_MS=30000
KNOWLEDGE_PDF_RASTERIZER_MAX_ASSETS=500
KNOWLEDGE_PDF_RASTERIZER_MAX_CONCURRENCY=2

# Ingestion-time LLM concurrency. The per-document outline limit improves one document's latency;
# the process-wide fair limit prevents simultaneous imports from multiplying provider pressure.
# The global value is per API replica and also covers embedding transport requests.
KNOWLEDGE_OUTLINE_SUMMARY_MAX_CONCURRENCY=8
KNOWLEDGE_OUTLINE_SUMMARY_BATCH_SIZE=8
KNOWLEDGE_OUTLINE_SUMMARY_BATCH_MAX_INPUT_CHARS=32000
KNOWLEDGE_MODEL_RUNTIME_GLOBAL_CONCURRENCY=16
KNOWLEDGE_SEMANTIC_EXTRACTION_BATCH_SIZE=8
KNOWLEDGE_SEMANTIC_EXTRACTION_MAX_CONCURRENCY=4
# Dify/plugin-daemon embedding requests carry at most 16 texts each. Two concurrent requests
# improve large-document latency while keeping high-dimensional response memory bounded.
KNOWLEDGE_EMBEDDING_REQUEST_CONCURRENCY=2
# Visual embeddings follow each knowledge space's active embedding capability. Keep each enabled
# image-byte request below both a count and raw-byte ceiling; these values never select a model.
KNOWLEDGE_VISUAL_EMBEDDING_MAX_ASSET_BYTES=20971520
KNOWLEDGE_VISUAL_EMBEDDING_MAX_BATCH_ASSETS=8
KNOWLEDGE_VISUAL_EMBEDDING_MAX_BATCH_BYTES=33554432
KNOWLEDGE_VISUAL_EMBEDDING_MAX_CONCURRENCY=2
# Document image understanding and answer image blocks follow the active reasoning capability.
# These ceilings bound per-process VLM concurrency and raw bytes retained before base64 encoding.
KNOWLEDGE_MULTIMODAL_ENRICHMENT_MAX_CONCURRENCY=4
KNOWLEDGE_MULTIMODAL_ENRICHMENT_MAX_IMAGE_BYTES=10485760
KNOWLEDGE_MULTIMODAL_ANSWER_MAX_CONCURRENCY=2
KNOWLEDGE_MULTIMODAL_ANSWER_MAX_IMAGE_ATTACHMENTS=8
KNOWLEDGE_MULTIMODAL_ANSWER_MAX_IMAGE_BYTES=10485760
KNOWLEDGE_MULTIMODAL_ANSWER_MAX_TOTAL_IMAGE_BYTES=33554432

# Required before enabling integrated product traffic. Set ENABLED=true only after providing the
# public-only current/previous JWKS that matches Dify's signer. Never copy Dify's private key here.
KNOWLEDGE_FS_CAPABILITY_V2_ENABLED=false
KNOWLEDGE_FS_CAPABILITY_V2_PUBLIC_JWKS=

# Internal Dify API-to-KnowledgeFS data paths. These flags do not expose browser CORS routes.
# Integrated document staging depends on upload sessions. Keep the default upload, multipart,
# and parser envelopes aligned with Dify's current 15 MiB document limit. Direct KnowledgeFS
# clients may raise the upload/parser limits together up to the documented 50 MiB hard ceiling;
# Dify product routes remain capped by their 15 MiB product-operation contract.
KNOWLEDGE_DIRECT_UPLOAD_ENABLED=on
KNOWLEDGE_DIRECT_UPLOAD_MAX_FILE_BYTES=15728640
KNOWLEDGE_DIRECT_UPLOAD_MULTIPART_THRESHOLD_BYTES=15728640
KNOWLEDGE_DIRECT_UPLOAD_SMALL_FALLBACK_MAX_BYTES=15728640
# The integrated Dify adapter cannot presign uploads, so this compatibility path buffers request
# bytes before writing them. Bound both active uploads and their aggregate retained payload bytes.
KNOWLEDGE_DIRECT_UPLOAD_SMALL_FALLBACK_MAX_CONCURRENCY=2
KNOWLEDGE_DIRECT_UPLOAD_SMALL_FALLBACK_MAX_RESERVED_BYTES=31457280
# Direct multipart document routes use a separate pre-validator gate. The retained-byte charge is
# conservative (bounded stream chunks + Hono caches + File.arrayBuffer), not an RSS measurement.
KNOWLEDGE_BUFFERED_DOCUMENT_UPLOAD_MAX_CONCURRENCY=2
KNOWLEDGE_BUFFERED_DOCUMENT_UPLOAD_MAX_RESERVED_BYTES=201326592
KNOWLEDGE_BUFFERED_DOCUMENT_UPLOAD_IDLE_TIMEOUT_MS=30000
KNOWLEDGE_BUFFERED_DOCUMENT_UPLOAD_TOTAL_TIMEOUT_MS=600000
# Enable Research task SSE only when the durable task progress repository is ready.
KNOWLEDGE_DIRECT_STREAM_ENABLED=off
# One generous cap avoids a second full Judge call when a reasoning model consumes hidden tokens.
# Judge prompts still require compact JSON and supported OpenAI reasoning models use low effort.
KNOWLEDGE_RESEARCH_REASONING_MAX_OUTPUT_TOKENS=8192
KNOWLEDGE_RESEARCH_REASONING_TIMEOUT_MS=60000
# One initial pool is shared fairly across the original query and all planned intents.
# Lower it to trade recall depth for latency. A durable supplemental list is separately plan-bounded.
KNOWLEDGE_RESEARCH_MAX_RERANK_CANDIDATES=200

# Query-image retrieval follows the active knowledge-space embedding/reasoning model capabilities.
# Deep/Research perform at most one vision expansion per query; durable Research retries reuse it.
KNOWLEDGE_QUERY_IMAGE_EXPANSION_TIMEOUT_MS=8000

# The bundled profile uses this internal service name. Override it with an external
# Unstructured-compatible endpoint when the deployment owns that provider instead.
UNSTRUCTURED_API_URL=http://knowledge_fs_unstructured:8000
UNSTRUCTURED_API_KEY=
# This is the process-wide ceiling shared by all remote parser requests. The nested heavy limit
# covers every PDF plus structurally/byte-heavy Office, email, EPUB, ODT, and RTF request.
UNSTRUCTURED_MAX_CONCURRENCY=2
UNSTRUCTURED_HEAVY_MAX_CONCURRENCY=1
# Keep the parser cap aligned with the default 15 MiB product upload limit. Operators increasing
# the upload cap must raise this value deliberately; the API rejects values above 50 MiB.
UNSTRUCTURED_MAX_INPUT_BYTES=15728640
# Ordinary documents use a generic ten-minute deadline. Only classified heavy documents get
# bounded extra headroom, so a small DOCX/PPTX/XLSX does not inherit the 40-minute ceiling.
UNSTRUCTURED_REQUEST_TIMEOUT_MS=600000
UNSTRUCTURED_HEAVY_REQUEST_TIMEOUT_MS=2400000
UNSTRUCTURED_MAX_RESPONSE_BYTES=33554432
# Durable compilation owns attempt retries. Do not multiply them with parser-client retries.
UNSTRUCTURED_MAX_RETRIES=0

# Bound authenticated Dify object-storage calls so cleanup cannot hang a compilation lease forever.
DIFY_OBJECT_STORAGE_REQUEST_TIMEOUT_MS=60000
DIFY_REMOTE_IMAGE_REQUEST_TIMEOUT_MS=30000