feat: playbook AI Action (#1693)

* feat: playbook AI Action

* feat: option to use agents

[skip ci]

* feat: ollama support

[skip ci]

* feat: support multiple relationships

* just provide a minimal graph without the config (manifests)

[skip ci]

* feat: make config optional. use the playbook's resource config.

* feat: better prompt

* feat: add system prompts

* feat: date support when fetching analysis

* chore: send knowledge base as JSON

* feat: connection support for LLMS

* chore: add fixture to send ai diagnosis to slack using BlockKit

* feat: add different output formats to the result

Makefile

@@ -93,7 +93,10 @@ gen-schemas:
 	cd hack/generate-schemas && \
 	go mod edit -module=github.com/flanksource/incident-commander/hack/generate-schemas && \
 	go mod edit -require=github.com/flanksource/[email protected] && \
- 	go mod edit -replace=github.com/flanksource/incident-commander=../../ && \
+	go mod edit -replace=github.com/flanksource/incident-commander=../../ && \
+	if grep -v "^//" ../../go.mod | grep -q "replace.*github.com/flanksource/duty.*=>"; then \
+		go mod edit -replace=github.com/flanksource/duty=../../../duty; \
+	fi && \
 	go mod tidy && \
 	go run ./main.go
 

api/global.go

@@ -26,3 +26,11 @@ var (
 const (
 	PropertyIncidentsDisabled = "incidents.disable"
 )
+
+type LLMBackend string
+
+const (
+	LLMBackendAnthropic LLMBackend = "anthropic"
+	LLMBackendOpenAI    LLMBackend = "openai"
+	LLMBackendOllama    LLMBackend = "ollama"
+)

api/v1/connection_types.go

@@ -238,15 +238,36 @@ type ConnectionSMB struct {
 	Share string `json:"share"`
 }
 
+type ConnectionOpenAI struct {
+	Model   *string       `json:"model,omitempty"`
+	BaseURL *types.EnvVar `json:"url,omitempty"`
+	ApiKey  types.EnvVar  `json:"apiKey"`
+}
+
+type ConnectionOllama struct {
+	Model   *string      `json:"model,omitempty"`
+	BaseURL types.EnvVar `json:"url,omitempty"`
+	ApiKey  types.EnvVar `json:"apiKey,omitempty"`
+}
+
+type ConnectionAnthropic struct {
+	Model   *string       `json:"model,omitempty"`
+	BaseURL *types.EnvVar `json:"url,omitempty"`
+	ApiKey  types.EnvVar  `json:"apiKey"`
+}
+
 // ConnectionSpec defines the desired state of Connection
 type ConnectionSpec struct {
 	Properties types.JSONStringMap `json:"properties,omitempty"`
 
+	Anthropic   *ConnectionAnthropic   `json:"anthropic,omitempty"`
 	AWS         *ConnectionAWS         `json:"aws,omitempty"`
-	S3          *ConnectionAWSS3       `json:"s3,omitempty"`
 	Azure       *ConnectionAzure       `json:"azure,omitempty"`
 	AzureDevops *ConnectionAzureDevops `json:"azureDevops,omitempty"`
 	GCP         *ConnectionGCP         `json:"gcp,omitempty"`
+	Ollama      *ConnectionOllama      `json:"ollama,omitempty"`
+	OpenAI      *ConnectionOpenAI      `json:"openai,omitempty"`
+	S3          *ConnectionAWSS3       `json:"s3,omitempty"`
 
 	Folder     *ConnectionFolder     `json:"folder,omitempty"`
 	Git        *ConnectionGit        `json:"git,omitempty"`

api/v1/playbook_actions.go

@@ -9,9 +9,15 @@ import (
 	"github.com/flanksource/commons/duration"
 	"github.com/flanksource/commons/utils"
 	"github.com/flanksource/duty/connection"
+	"github.com/flanksource/duty/context"
 	"github.com/flanksource/duty/models"
+	"github.com/flanksource/duty/query"
 	"github.com/flanksource/duty/types"
+	"github.com/google/uuid"
+	"github.com/samber/lo"
 	"k8s.io/client-go/kubernetes"
+
+	"github.com/flanksource/incident-commander/api"
 )
 
 type NotificationAction struct {
@@ -235,6 +241,157 @@ func (git GitCheckout) GetCertificate() types.EnvVar {
 	return utils.Deref(git.Certificate)
 }
 
+type TimeMetadata struct {
+	Since string `json:"since" yaml:"since"`
+}
+
+type AIActionRelationship struct {
+	// max depth to traverse the relationship. Defaults to 3
+	Depth *int `json:"depth,omitempty"`
+
+	// use incoming/outgoing/all relationships.
+	Direction query.RelationDirection `json:"direction,omitempty"`
+
+	Changes  TimeMetadata `json:"changes,omitempty"`
+	Analysis TimeMetadata `json:"analysis,omitempty"`
+}
+
+func (t AIActionRelationship) ToRelationshipQuery(configID uuid.UUID) query.RelationQuery {
+	q := query.RelationQuery{
+		ID:       configID,
+		MaxDepth: t.Depth,
+		Relation: t.Direction,
+	}
+
+	if q.MaxDepth == nil {
+		q.MaxDepth = lo.ToPtr(3)
+	}
+
+	if q.Relation == "" {
+		q.Relation = query.All
+	}
+
+	return q
+}
+
+type AIActionClient struct {
+	// Connection to setup the llm backend connection
+	Connection *string `json:"connection,omitempty"`
+
+	// API Key
+	APIKey types.EnvVar `json:"apiKey,omitempty"`
+
+	// Optionally specify the LLM backend.
+	// Supported: anthropic (default), ollama, openai.
+	Backend api.LLMBackend `json:"backend,omitempty"`
+
+	// Model name based on the backend chosen.
+	// Example: gpt-4o for openai, claude-3-5-sonnet-latest for Anthropic, llama3.1:8b for Ollama
+	Model string `json:"model,omitempty"`
+
+	// BaseURL or API url.
+	// Example: server URL for ollama or custom url for Anthropic if using a proxy
+	APIURL string `json:"apiURL,omitempty"`
+}
+
+func (t *AIActionClient) Populate(ctx context.Context) error {
+	if t.Connection != nil {
+		conn, err := ctx.HydrateConnectionByURL(*t.Connection)
+		if err != nil {
+			return err
+		} else if conn == nil {
+			return fmt.Errorf("connection(%s) was not found: %w", *t.Connection, err)
+		}
+
+		if err := t.APIKey.Scan(conn.Password); err != nil {
+			return err
+		}
+
+		t.APIURL = conn.URL
+
+		if m, ok := conn.Properties["model"]; ok {
+			t.Model = m
+		}
+
+		switch conn.Type {
+		case models.ConnectionTypeOllama:
+			t.Backend = api.LLMBackendOllama
+		case models.ConnectionTypeAnthropic:
+			t.Backend = api.LLMBackendAnthropic
+		case models.ConnectionTypeOpenAI:
+			t.Backend = api.LLMBackendOpenAI
+		default:
+			return fmt.Errorf("connection of type %q is not supported. Supported types: %s, %s & %s",
+				conn.Type, models.ConnectionTypeOllama, models.ConnectionTypeAnthropic, models.ConnectionTypeOpenAI,
+			)
+		}
+	}
+
+	if !t.APIKey.IsEmpty() {
+		if v, err := ctx.GetEnvValueFromCache(t.APIKey, ctx.GetNamespace()); err != nil {
+			return fmt.Errorf("failed to get api key from source ref : %w", err)
+		} else {
+			t.APIKey.ValueStatic = v
+		}
+	}
+
+	return nil
+}
+
+type AIActionContext struct {
+	// The config id to operate on.
+	// If not provided, the playbook's config is used.
+	Config string `json:"config,omitempty" yaml:"config,omitempty" template:"true"`
+
+	// Select changes for the config to provide as an additional context to the AI model.
+	Changes TimeMetadata `json:"changes,omitempty" yaml:"changes,omitempty"`
+
+	// Select analysis for the config to provide as an additional context to the AI model.
+	Analysis TimeMetadata `json:"analysis,omitempty" yaml:"analysis,omitempty"`
+
+	// Select related configs to provide as an additional context to the AI model.
+	Relationships []AIActionRelationship `json:"relationships,omitempty" yaml:"relationships,omitempty"`
+}
+
+func (t AIActionContext) ShouldFetchConfigChanges() bool {
+	// if changes are being fetched from relationships, we don't have to query
+	// the changes for just the config alone.
+
+	if t.Changes.Since == "" {
+		return false
+	}
+
+	for _, r := range t.Relationships {
+		if r.Changes.Since != "" {
+			return false
+		}
+	}
+
+	return true
+}
+
+type AIAction struct {
+	AIActionClient  `json:",inline" yaml:",inline"`
+	AIActionContext `json:",inline" yaml:",inline" template:"true"`
+
+	// Use an AI agent that can autonomously drive the diagnosis using tools that interface directly with the database.
+	// NOTE: Not exposed for now
+	UseAgent bool `json:"-"`
+
+	// system prompt is a way to provide context, instructions, and guidelines to the LLM before presenting it
+	// with a question or task.
+	// By using a system prompt, you can set the stage for the conversation, specifying LLM's role, personality,
+	// tone, or any other relevant information that will help it better understand and respond to the user's input.
+	SystemPrompt string `json:"systemPrompt"`
+
+	// Prompt is the humna prompt
+	Prompt string `json:"prompt" template:"true"`
+
+	// Output format of the prompt.
+	// Supported: markdown (default), slack.
+	Formats []string `json:"formats,omitempty"`
+}
+
 type ExecAction struct {
 	// Script can be an inline script or a path to a script that needs to be executed
 	// On windows executed via powershell and in darwin and linux executed using bash
@@ -405,6 +562,7 @@ type PlaybookAction struct {
 	// When left empty, the templating is done on the main instance(host) itself.
 	TemplatesOn string `json:"templatesOn,omitempty" yaml:"templatesOn,omitempty"`
 
+	AI                  *AIAction                  `json:"ai,omitempty" yaml:"ai,omitempty" template:"true"`
 	Exec                *ExecAction                `json:"exec,omitempty" yaml:"exec,omitempty" template:"true"`
 	GitOps              *GitOpsAction              `json:"gitops,omitempty" yaml:"gitops,omitempty" template:"true"`
 	Github              *GithubAction              `json:"github,omitempty" yaml:"github,omitempty" template:"true"`