thinkwork-cli 0.12.12 → 0.12.15

package/dist/plugins/company-brain/terraform/cognee/main.tf

@@ -0,0 +1,766 @@
+################################################################################
+# Cognee (Optional Ontology/KG Add-On) - App Module
+#
+# Provisions the phase-1 Cognee substrate as an internal ECS/Fargate service.
+# The module intentionally does not expose a public endpoint mode; Thinkwork API
+# wrappers own tenant/auth boundaries in follow-up work.
+################################################################################
+
+locals {
+  legacy_name         = "thinkwork-${var.stage}-cognee"
+  legacy_cluster_name = "thinkwork-${var.stage}-brain-cluster"
+
+  raw_brain_instance_key        = var.brain_instance_key != "" ? var.brain_instance_key : var.brain_tenant_id
+  normalized_brain_instance_key = trim(replace(replace(lower(local.raw_brain_instance_key), "/[^a-z0-9-]/", "-"), "/-+/", "-"), "-")
+  tenant_scoped_brain_instance  = local.normalized_brain_instance_key != ""
+  brain_instance_hash           = substr(sha1(local.raw_brain_instance_key), 0, 12)
+  name                          = local.tenant_scoped_brain_instance ? "thinkwork-${var.stage}-cb-${local.brain_instance_hash}" : local.legacy_name
+  cluster_name                  = local.tenant_scoped_brain_instance ? "${local.name}-cluster" : local.legacy_cluster_name
+  resource_short_name           = local.tenant_scoped_brain_instance ? "tw-${substr(var.stage, 0, 8)}-cb-${substr(local.brain_instance_hash, 0, 10)}" : "tw-${var.stage}-cognee"
+
+  data_root_directory   = "/app/cognee-storage/data"
+  system_root_directory = "/app/cognee-storage/system"
+
+  vector_db_url = (
+    var.vector_db_url != ""
+    ? var.vector_db_url
+    : var.vector_db_provider == "neptune_analytics" && var.neptune_endpoint != ""
+    ? var.neptune_endpoint
+    : "${local.system_root_directory}/databases/cognee.lancedb"
+  )
+  graph_db_url = (
+    var.graph_database_url != ""
+    ? var.graph_database_url
+    : var.graph_database_provider == "neptune_analytics" && var.neptune_endpoint != ""
+    ? var.neptune_endpoint
+    : "${local.system_root_directory}/databases/cognee.kuzu"
+  )
+  log_group_name = local.tenant_scoped_brain_instance ? "/thinkwork/${var.stage}/brain/${local.normalized_brain_instance_key}/cognee" : "/thinkwork/${var.stage}/cognee"
+
+  managed_secret_specs = {
+    db_password = {
+      enabled       = var.create_secret_placeholders && var.db_password_secret_arn == ""
+      name          = local.tenant_scoped_brain_instance ? "thinkwork/${var.stage}/brain/${local.normalized_brain_instance_key}/db-credentials" : "thinkwork/${var.stage}/cognee/db-credentials"
+      description   = "Dedicated Brain/Cognee PostgreSQL credentials"
+      secret_string = jsonencode({ password = "PLACEHOLDER_SET_VIA_CLI" })
+    }
+    llm_api_key = {
+      enabled       = var.create_secret_placeholders && var.llm_provider != "bedrock" && var.llm_api_key_secret_arn == ""
+      name          = local.tenant_scoped_brain_instance ? "thinkwork/${var.stage}/brain/${local.normalized_brain_instance_key}/llm-api-key" : "thinkwork/${var.stage}/cognee/llm-api-key"
+      description   = "Brain/Cognee non-Bedrock LLM provider API key"
+      secret_string = "PLACEHOLDER_SET_VIA_CLI"
+    }
+    embedding_api_key = {
+      enabled       = var.create_secret_placeholders && var.embedding_provider != "bedrock" && var.embedding_api_key_secret_arn == ""
+      name          = local.tenant_scoped_brain_instance ? "thinkwork/${var.stage}/brain/${local.normalized_brain_instance_key}/embedding-api-key" : "thinkwork/${var.stage}/cognee/embedding-api-key"
+      description   = "Brain/Cognee non-Bedrock embedding provider API key"
+      secret_string = "PLACEHOLDER_SET_VIA_CLI"
+    }
+    vector_db_key = {
+      enabled       = var.create_secret_placeholders && var.vector_db_provider != "lancedb" && var.vector_db_key_secret_arn == ""
+      name          = local.tenant_scoped_brain_instance ? "thinkwork/${var.stage}/brain/${local.normalized_brain_instance_key}/vector-db-key" : "thinkwork/${var.stage}/cognee/vector-db-key"
+      description   = "Brain/Cognee remote vector store key"
+      secret_string = "PLACEHOLDER_SET_VIA_CLI"
+    }
+    graph_database_password = {
+      enabled       = var.create_secret_placeholders && var.graph_database_provider != "kuzu" && var.graph_database_password_secret_arn == ""
+      name          = local.tenant_scoped_brain_instance ? "thinkwork/${var.stage}/brain/${local.normalized_brain_instance_key}/graph-database-password" : "thinkwork/${var.stage}/cognee/graph-database-password"
+      description   = "Brain/Cognee remote graph store password"
+      secret_string = "PLACEHOLDER_SET_VIA_CLI"
+    }
+  }
+
+  managed_secrets = {
+    for key, spec in local.managed_secret_specs : key => spec if spec.enabled
+  }
+
+  effective_db_password_secret_arn = (
+    var.db_password_secret_arn != ""
+    ? var.db_password_secret_arn
+    : try(aws_secretsmanager_secret.cognee["db_password"].arn, "")
+  )
+  effective_llm_api_key_secret_arn = (
+    var.llm_api_key_secret_arn != ""
+    ? var.llm_api_key_secret_arn
+    : try(aws_secretsmanager_secret.cognee["llm_api_key"].arn, "")
+  )
+  effective_embedding_api_key_secret_arn = (
+    var.embedding_api_key_secret_arn != ""
+    ? var.embedding_api_key_secret_arn
+    : try(aws_secretsmanager_secret.cognee["embedding_api_key"].arn, "")
+  )
+  effective_vector_db_key_secret_arn = (
+    var.vector_db_key_secret_arn != ""
+    ? var.vector_db_key_secret_arn
+    : try(aws_secretsmanager_secret.cognee["vector_db_key"].arn, "")
+  )
+  effective_graph_database_password_secret_arn = (
+    var.graph_database_password_secret_arn != ""
+    ? var.graph_database_password_secret_arn
+    : try(aws_secretsmanager_secret.cognee["graph_database_password"].arn, "")
+  )
+
+  efs_mount_subnet_ids_by_index = {
+    for index, subnet_id in var.subnet_ids : tostring(index) => subnet_id
+  }
+
+  brain_artifact_object_arns = [
+    for prefix in var.brain_artifacts_prefixes :
+    "${var.brain_artifacts_bucket_arn}/${trim(prefix, "/")}/*"
+  ]
+
+  base_environment = [
+    { name = "DB_PROVIDER", value = "postgres" },
+    { name = "DB_HOST", value = var.db_host },
+    { name = "DB_PORT", value = tostring(var.db_port) },
+    { name = "DB_NAME", value = var.db_name },
+    { name = "DB_USERNAME", value = var.db_username },
+    { name = "LLM_PROVIDER", value = var.llm_provider },
+    { name = "LLM_MODEL", value = var.llm_model },
+    { name = "EMBEDDING_PROVIDER", value = var.embedding_provider },
+    { name = "EMBEDDING_MODEL", value = var.embedding_model },
+    { name = "EMBEDDING_DIMENSIONS", value = tostring(var.embedding_dimensions) },
+    { name = "VECTOR_DB_PROVIDER", value = var.vector_db_provider },
+    { name = "VECTOR_DB_URL", value = local.vector_db_url },
+    { name = "GRAPH_DATABASE_PROVIDER", value = var.graph_database_provider },
+    { name = "GRAPH_DATABASE_URL", value = local.graph_db_url },
+    { name = "DATA_ROOT_DIRECTORY", value = local.data_root_directory },
+    { name = "SYSTEM_ROOT_DIRECTORY", value = local.system_root_directory },
+    { name = "BRAIN_TENANT_ID", value = var.brain_tenant_id },
+    { name = "BRAIN_INSTANCE_KEY", value = local.normalized_brain_instance_key },
+    { name = "BRAIN_STORAGE_TIER", value = var.brain_storage_tier },
+    { name = "BRAIN_S3_ARTIFACT_ROOT", value = var.brain_s3_artifact_root },
+    { name = "BRAIN_S3_MANIFEST_ROOT", value = var.brain_s3_manifest_root },
+    { name = "BRAIN_S3_VAULT_PROJECTION_ROOT", value = var.brain_s3_vault_projection_root },
+    { name = "NEPTUNE_GRAPH_ID", value = var.neptune_graph_id },
+    { name = "NEPTUNE_ENDPOINT", value = var.neptune_endpoint },
+    { name = "TELEMETRY_DISABLED", value = "true" },
+    { name = "REQUIRE_AUTHENTICATION", value = tostring(var.require_authentication) },
+    { name = "ENABLE_BACKEND_ACCESS_CONTROL", value = tostring(var.enable_backend_access_control) },
+    { name = "CORS_ALLOWED_ORIGINS", value = var.cors_allowed_origins },
+    { name = "AWS_DEFAULT_REGION", value = data.aws_region.current.name },
+    { name = "AWS_REGION", value = data.aws_region.current.name },
+    # litellm's default embedding timeout (30s) is too tight for a batch of
+    # entity embeddings routed sequentially through the Bedrock VPC endpoint —
+    # cognify errored with EmbeddingException (timeout) mid-run. Raise the
+    # per-request embedding + LLM timeouts so a batch completes.
+    { name = "EMBEDDING_REQUEST_TIMEOUT", value = "180" },
+    { name = "LLM_REQUEST_TIMEOUT", value = "180" },
+    # Bedrock throttles (HTTP 429) the burst of per-entity titan embedding
+    # calls cognify fires — the pipeline errored on EmbeddingException. The
+    # account's hard quota is exactly 60 req/min for Titan Text Embeddings V2
+    # (Service Quotas), so a limiter set AT 60 still tips over under any
+    # concurrency. Pace embeddings to 25/min — comfortably under the ceiling
+    # with headroom for litellm retries. The single dogfood task is slow but
+    # completes; the self-invoke ingest chain drains the backlog across runs.
+    # Production-scale volume needs a Titan-embed quota increase (Service
+    # Quotas) — 60 RPM caps throughput regardless of pacing.
+    { name = "EMBEDDING_RATE_LIMIT_ENABLED", value = "true" },
+    { name = "EMBEDDING_RATE_LIMIT_REQUESTS", value = "25" },
+    { name = "EMBEDDING_RATE_LIMIT_INTERVAL", value = "60" },
+    # Cognee's generic LLM rate limiter (shared limiter implementation) —
+    # enable it too so the Kimi extraction calls don't trip Bedrock either.
+    { name = "LLM_RATE_LIMIT_ENABLED", value = "true" },
+    { name = "LLM_RATE_LIMIT_REQUESTS", value = "30" },
+    { name = "LLM_RATE_LIMIT_INTERVAL", value = "60" },
+  ]
+
+  optional_environment = concat(
+    var.embedding_max_completion_tokens != "" ? [{ name = "EMBEDDING_MAX_COMPLETION_TOKENS", value = var.embedding_max_completion_tokens }] : [],
+    var.graph_database_username != "" ? [{ name = "GRAPH_DATABASE_USERNAME", value = var.graph_database_username }] : []
+  )
+
+  container_secrets = concat(
+    [{ name = "DB_PASSWORD", valueFrom = "${local.effective_db_password_secret_arn}:password::" }],
+    local.effective_llm_api_key_secret_arn != "" ? [{ name = "LLM_API_KEY", valueFrom = local.effective_llm_api_key_secret_arn }] : [],
+    local.effective_embedding_api_key_secret_arn != "" ? [{ name = "EMBEDDING_API_KEY", valueFrom = local.effective_embedding_api_key_secret_arn }] : [],
+    local.effective_vector_db_key_secret_arn != "" ? [{ name = "VECTOR_DB_KEY", valueFrom = local.effective_vector_db_key_secret_arn }] : [],
+    local.effective_graph_database_password_secret_arn != "" ? [{ name = "GRAPH_DATABASE_PASSWORD", valueFrom = local.effective_graph_database_password_secret_arn }] : []
+  )
+
+  secret_arns = compact([
+    local.effective_db_password_secret_arn,
+    local.effective_llm_api_key_secret_arn,
+    local.effective_embedding_api_key_secret_arn,
+    local.effective_vector_db_key_secret_arn,
+    local.effective_graph_database_password_secret_arn,
+  ])
+}
+
+data "aws_region" "current" {}
+
+resource "aws_secretsmanager_secret" "cognee" {
+  for_each = local.managed_secrets
+
+  name        = each.value.name
+  description = each.value.description
+
+  tags = {
+    Name = each.value.name
+    Role = "cognee"
+  }
+}
+
+resource "aws_secretsmanager_secret_version" "cognee" {
+  for_each = local.managed_secrets
+
+  secret_id     = aws_secretsmanager_secret.cognee[each.key].id
+  secret_string = each.value.secret_string
+
+  lifecycle {
+    ignore_changes = [secret_string]
+  }
+}
+
+################################################################################
+# Plan-time safety guardrails
+################################################################################
+
+resource "terraform_data" "configuration_guardrails" {
+  input = {
+    backend_mode                 = var.backend_mode
+    brain_tenant_id              = var.brain_tenant_id
+    brain_instance_key           = local.normalized_brain_instance_key
+    brain_storage_tier           = var.brain_storage_tier
+    desired_count                = var.desired_count
+    vector_db_url                = local.vector_db_url
+    vector_db_provider           = var.vector_db_provider
+    graph_database_url           = local.graph_db_url
+    graph_database_provider      = var.graph_database_provider
+    neptune_graph_id             = var.neptune_graph_id
+    neptune_endpoint             = var.neptune_endpoint
+    llm_provider                 = var.llm_provider
+    embedding_provider           = var.embedding_provider
+    db_password_secret_arn       = local.effective_db_password_secret_arn
+    llm_api_key_secret_arn       = local.effective_llm_api_key_secret_arn
+    embedding_api_key_secret_arn = local.effective_embedding_api_key_secret_arn
+    bedrock_model_resource_arns  = var.bedrock_model_resource_arns
+  }
+
+  lifecycle {
+    precondition {
+      condition     = var.backend_mode != "dogfood" || var.desired_count == 1
+      error_message = "dogfood backend mode uses local graph/vector storage and must run with desired_count = 1."
+    }
+
+    precondition {
+      condition = (
+        var.brain_storage_tier != "default" ||
+        (
+          var.backend_mode == "dogfood" &&
+          var.vector_db_provider == "lancedb" &&
+          var.graph_database_provider == "kuzu" &&
+          var.desired_count == 1
+        )
+      )
+      error_message = "default Brain tier must use dogfood backend mode, lancedb vectors, kuzu graph storage, and desired_count = 1."
+    }
+
+    precondition {
+      condition = (
+        var.brain_storage_tier != "production" ||
+        (
+          var.backend_mode == "remote" &&
+          var.vector_db_provider == "neptune_analytics" &&
+          var.graph_database_provider == "neptune_analytics" &&
+          var.neptune_graph_id != "" &&
+          var.neptune_endpoint != ""
+        )
+      )
+      error_message = "production Brain tier must use remote mode with Neptune Analytics graph/vector providers, neptune_graph_id, and neptune_endpoint."
+    }
+
+    precondition {
+      condition     = !contains(["opensearch", "opensearch_serverless", "aoss"], lower(var.vector_db_provider))
+      error_message = "Company Brain production must not use direct OpenSearch vector storage; use Neptune Analytics for production graph/vector."
+    }
+
+    precondition {
+      condition     = var.private_substrate_mode
+      error_message = "Company Brain substrate must remain private/internal-only."
+    }
+
+    precondition {
+      condition = (
+        var.backend_mode != "remote" ||
+        (
+          local.vector_db_url != "" &&
+          local.graph_db_url != "" &&
+          local.vector_db_url != "${local.system_root_directory}/databases/cognee.lancedb" &&
+          local.graph_db_url != "${local.system_root_directory}/databases/cognee.kuzu"
+        )
+      )
+      error_message = "remote backend mode requires vector and graph URLs, or Neptune Analytics providers with neptune_endpoint."
+    }
+
+    precondition {
+      condition     = local.effective_db_password_secret_arn != ""
+      error_message = "Cognee requires db_password_secret_arn or create_secret_placeholders = true."
+    }
+
+    precondition {
+      condition = (
+        var.llm_provider == "bedrock" || local.effective_llm_api_key_secret_arn != ""
+        ) && (
+        var.embedding_provider == "bedrock" || local.effective_embedding_api_key_secret_arn != ""
+      )
+      error_message = "Non-Bedrock LLM or embedding providers must use Secrets Manager ARNs, not plaintext environment values."
+    }
+
+    precondition {
+      condition = (
+        (var.llm_provider != "bedrock" && var.embedding_provider != "bedrock") ||
+        length(var.bedrock_model_resource_arns) > 0
+      )
+      error_message = "Bedrock LLM or embedding providers require explicit bedrock_model_resource_arns."
+    }
+  }
+}
+
+################################################################################
+# ECS Cluster
+################################################################################
+
+resource "aws_ecs_cluster" "main" {
+  name = local.cluster_name
+
+  setting {
+    name  = "containerInsights"
+    value = "enabled"
+  }
+
+  tags = { Name = local.cluster_name }
+}
+
+################################################################################
+# IAM
+################################################################################
+
+resource "aws_iam_role" "ecs_execution" {
+  name = "${local.name}-execution"
+
+  assume_role_policy = jsonencode({
+    Version = "2012-10-17"
+    Statement = [{
+      Action    = "sts:AssumeRole"
+      Effect    = "Allow"
+      Principal = { Service = "ecs-tasks.amazonaws.com" }
+    }]
+  })
+}
+
+resource "aws_iam_role_policy_attachment" "ecs_execution" {
+  role       = aws_iam_role.ecs_execution.name
+  policy_arn = "arn:aws:iam::aws:policy/service-role/AmazonECSTaskExecutionRolePolicy"
+}
+
+resource "aws_iam_role_policy" "ecs_execution_secrets" {
+  name = "cognee-secrets"
+  role = aws_iam_role.ecs_execution.id
+
+  policy = jsonencode({
+    Version = "2012-10-17"
+    Statement = concat(
+      [{
+        Effect   = "Allow"
+        Action   = ["secretsmanager:GetSecretValue"]
+        Resource = local.secret_arns
+      }],
+      length(var.kms_key_arns) > 0 ? [{
+        Effect   = "Allow"
+        Action   = ["kms:Decrypt"]
+        Resource = var.kms_key_arns
+      }] : []
+    )
+  })
+}
+
+resource "aws_iam_role" "ecs_task" {
+  name = "${local.name}-task"
+
+  assume_role_policy = jsonencode({
+    Version = "2012-10-17"
+    Statement = [{
+      Action    = "sts:AssumeRole"
+      Effect    = "Allow"
+      Principal = { Service = "ecs-tasks.amazonaws.com" }
+    }]
+  })
+}
+
+resource "aws_iam_role_policy" "bedrock_access" {
+  count = var.llm_provider == "bedrock" || var.embedding_provider == "bedrock" ? 1 : 0
+
+  name = "bedrock-access"
+  role = aws_iam_role.ecs_task.id
+
+  policy = jsonencode({
+    Version = "2012-10-17"
+    Statement = [{
+      Effect   = "Allow"
+      Action   = ["bedrock:InvokeModel", "bedrock:InvokeModelWithResponseStream"]
+      Resource = var.bedrock_model_resource_arns
+    }]
+  })
+}
+
+resource "aws_iam_role_policy" "brain_artifacts" {
+  count = var.brain_artifacts_bucket_arn != "" && length(var.brain_artifacts_prefixes) > 0 ? 1 : 0
+
+  name = "brain-artifacts"
+  role = aws_iam_role.ecs_task.id
+
+  policy = jsonencode({
+    Version = "2012-10-17"
+    Statement = [
+      {
+        Effect   = "Allow"
+        Action   = ["s3:ListBucket"]
+        Resource = var.brain_artifacts_bucket_arn
+        Condition = {
+          StringLike = {
+            "s3:prefix" = [for prefix in var.brain_artifacts_prefixes : "${trim(prefix, "/")}/*"]
+          }
+        }
+      },
+      {
+        Effect = "Allow"
+        Action = [
+          "s3:GetObject",
+          "s3:PutObject",
+          "s3:DeleteObject",
+          "s3:GetObjectVersion",
+        ]
+        Resource = local.brain_artifact_object_arns
+      },
+    ]
+  })
+}
+
+resource "aws_iam_role_policy" "neptune_graph_access" {
+  count = var.neptune_graph_arn != "" ? 1 : 0
+
+  name = "neptune-graph-access"
+  role = aws_iam_role.ecs_task.id
+
+  policy = jsonencode({
+    Version = "2012-10-17"
+    Statement = [{
+      Effect = "Allow"
+      Action = [
+        "neptune-graph:GetGraph",
+        "neptune-graph:ReadDataViaQuery",
+        "neptune-graph:WriteDataViaQuery",
+      ]
+      Resource = var.neptune_graph_arn
+    }]
+  })
+}
+
+# ECS Exec (`aws ecs execute-command`) needs the SSM messages channel on the
+# TASK role. Gated by var.enable_execute_command so it only grants when exec is
+# on. Used operationally to inspect the dogfood Cognee store / API directly
+# (its ALB is VPC-internal) — e.g. verify dataset names and graph-fetch scope.
+resource "aws_iam_role_policy" "ecs_exec_ssm" {
+  count = var.enable_execute_command ? 1 : 0
+
+  name = "ecs-exec-ssm-messages"
+  role = aws_iam_role.ecs_task.id
+
+  policy = jsonencode({
+    Version = "2012-10-17"
+    Statement = [{
+      Effect = "Allow"
+      Action = [
+        "ssmmessages:CreateControlChannel",
+        "ssmmessages:CreateDataChannel",
+        "ssmmessages:OpenControlChannel",
+        "ssmmessages:OpenDataChannel",
+      ]
+      Resource = "*"
+    }]
+  })
+}
+
+################################################################################
+# Security Groups
+################################################################################
+
+resource "aws_security_group" "cognee" {
+  name_prefix = "${local.name}-"
+  description = "Cognee ECS task"
+  vpc_id      = var.vpc_id
+
+  ingress {
+    description     = "Internal ALB to Cognee API"
+    from_port       = var.container_port
+    to_port         = var.container_port
+    protocol        = "tcp"
+    security_groups = [aws_security_group.alb.id]
+  }
+
+  egress {
+    from_port   = 0
+    to_port     = 0
+    protocol    = "-1"
+    cidr_blocks = ["0.0.0.0/0"]
+  }
+
+  tags = { Name = "${local.name}-sg" }
+  lifecycle { create_before_destroy = true }
+}
+
+resource "aws_security_group" "alb" {
+  name_prefix = "${local.name}-alb-"
+  description = "Internal Cognee ALB"
+  vpc_id      = var.vpc_id
+
+  dynamic "ingress" {
+    for_each = var.allowed_internal_cidr_blocks
+
+    content {
+      description = "Allowed internal caller CIDR"
+      from_port   = 80
+      to_port     = 80
+      protocol    = "tcp"
+      cidr_blocks = [ingress.value]
+    }
+  }
+
+  dynamic "ingress" {
+    for_each = var.allowed_internal_security_group_ids
+
+    content {
+      description     = "Allowed internal caller security group"
+      from_port       = 80
+      to_port         = 80
+      protocol        = "tcp"
+      security_groups = [ingress.value]
+    }
+  }
+
+  egress {
+    from_port   = 0
+    to_port     = 0
+    protocol    = "-1"
+    cidr_blocks = ["0.0.0.0/0"]
+  }
+
+  tags = { Name = "${local.name}-alb-sg" }
+  lifecycle { create_before_destroy = true }
+}
+
+resource "aws_security_group" "efs" {
+  name_prefix = "${local.name}-efs-"
+  description = "Cognee EFS storage"
+  vpc_id      = var.vpc_id
+
+  ingress {
+    description     = "Cognee task to EFS"
+    from_port       = 2049
+    to_port         = 2049
+    protocol        = "tcp"
+    security_groups = [aws_security_group.cognee.id]
+  }
+
+  egress {
+    from_port   = 0
+    to_port     = 0
+    protocol    = "-1"
+    cidr_blocks = ["0.0.0.0/0"]
+  }
+
+  tags = { Name = "${local.name}-efs-sg" }
+  lifecycle { create_before_destroy = true }
+}
+
+resource "aws_security_group_rule" "aurora_from_cognee" {
+  type                     = "ingress"
+  from_port                = var.db_port
+  to_port                  = var.db_port
+  protocol                 = "tcp"
+  source_security_group_id = aws_security_group.cognee.id
+  security_group_id        = var.db_security_group_id
+}
+
+################################################################################
+# ALB
+################################################################################
+
+resource "aws_lb" "cognee" {
+  name               = local.resource_short_name
+  internal           = true
+  load_balancer_type = "application"
+  security_groups    = [aws_security_group.alb.id]
+  subnets            = var.subnet_ids
+
+  tags = { Name = "${local.name}-alb" }
+}
+
+resource "aws_lb_target_group" "cognee" {
+  name        = local.resource_short_name
+  port        = var.container_port
+  protocol    = "HTTP"
+  vpc_id      = var.vpc_id
+  target_type = "ip"
+
+  health_check {
+    path                = var.health_check_path
+    port                = tostring(var.container_port)
+    protocol            = "HTTP"
+    healthy_threshold   = 2
+    unhealthy_threshold = 3
+    timeout             = 10
+    interval            = 30
+    matcher             = "200-399"
+  }
+
+  tags = { Name = "${local.name}-tg" }
+}
+
+resource "aws_lb_listener" "cognee" {
+  load_balancer_arn = aws_lb.cognee.arn
+  port              = 80
+  protocol          = "HTTP"
+
+  default_action {
+    type             = "forward"
+    target_group_arn = aws_lb_target_group.cognee.arn
+  }
+}
+
+################################################################################
+# Persistent writable storage
+################################################################################
+
+resource "aws_efs_file_system" "cognee" {
+  creation_token = local.name
+  encrypted      = true
+
+  tags = { Name = "${local.name}-efs" }
+}
+
+resource "aws_efs_mount_target" "cognee" {
+  for_each = local.efs_mount_subnet_ids_by_index
+
+  file_system_id  = aws_efs_file_system.cognee.id
+  subnet_id       = each.value
+  security_groups = [aws_security_group.efs.id]
+}
+
+################################################################################
+# CloudWatch Logs
+################################################################################
+
+resource "aws_cloudwatch_log_group" "cognee" {
+  name              = local.log_group_name
+  retention_in_days = var.log_retention_days
+
+  tags = { Name = "${local.name}-logs" }
+}
+
+################################################################################
+# ECS Task Definition
+################################################################################
+
+resource "aws_ecs_task_definition" "cognee" {
+  family                   = local.name
+  requires_compatibilities = ["FARGATE"]
+  network_mode             = "awsvpc"
+  cpu                      = var.cpu
+  memory                   = var.memory
+  execution_role_arn       = aws_iam_role.ecs_execution.arn
+  task_role_arn            = aws_iam_role.ecs_task.arn
+
+  runtime_platform {
+    operating_system_family = "LINUX"
+    cpu_architecture        = var.cpu_architecture
+  }
+
+  volume {
+    name = "cognee-storage"
+
+    efs_volume_configuration {
+      file_system_id     = aws_efs_file_system.cognee.id
+      transit_encryption = "ENABLED"
+    }
+  }
+
+  container_definitions = jsonencode([{
+    name      = "cognee"
+    image     = var.image_uri
+    essential = true
+
+    portMappings = [{
+      containerPort = var.container_port
+      protocol      = "tcp"
+    }]
+
+    environment = concat(local.base_environment, local.optional_environment)
+    secrets     = local.container_secrets
+
+    mountPoints = [{
+      sourceVolume  = "cognee-storage"
+      containerPath = "/app/cognee-storage"
+      readOnly      = false
+    }]
+
+    logConfiguration = {
+      logDriver = "awslogs"
+      options = {
+        "awslogs-group"         = aws_cloudwatch_log_group.cognee.name
+        "awslogs-region"        = data.aws_region.current.name
+        "awslogs-stream-prefix" = "cognee"
+      }
+    }
+
+    healthCheck = {
+      command     = ["CMD-SHELL", "curl -f http://localhost:${var.container_port}${var.health_check_path} || exit 1"]
+      interval    = 30
+      timeout     = 10
+      retries     = 3
+      startPeriod = 300
+    }
+  }])
+
+  depends_on = [aws_efs_mount_target.cognee, terraform_data.configuration_guardrails]
+
+  tags = { Name = local.name }
+}
+
+################################################################################
+# ECS Service
+################################################################################
+
+resource "aws_ecs_service" "cognee" {
+  name            = local.name
+  cluster         = aws_ecs_cluster.main.id
+  task_definition = aws_ecs_task_definition.cognee.arn
+  desired_count   = var.desired_count
+  launch_type     = "FARGATE"
+
+  # Allow `aws ecs execute-command` into the running container (operational
+  # introspection of the dogfood Cognee store/API, whose ALB is VPC-internal).
+  enable_execute_command = var.enable_execute_command
+
+  deployment_minimum_healthy_percent = var.backend_mode == "dogfood" ? 0 : 100
+  deployment_maximum_percent         = var.backend_mode == "dogfood" ? 100 : 200
+  health_check_grace_period_seconds  = var.health_check_grace_period_seconds
+  wait_for_steady_state              = var.wait_for_steady_state
+
+  network_configuration {
+    subnets          = var.subnet_ids
+    security_groups  = [aws_security_group.cognee.id]
+    assign_public_ip = true
+  }
+
+  load_balancer {
+    target_group_arn = aws_lb_target_group.cognee.arn
+    container_name   = "cognee"
+    container_port   = var.container_port
+  }
+
+  depends_on = [aws_lb_listener.cognee, terraform_data.configuration_guardrails]
+
+  tags = { Name = local.name }
+}