homelabby/internal/research/agent.go

package research

import (
	"context"
	"encoding/json"
	"fmt"
	"log"
	"regexp"
	"strings"
	"time"

	"git.georgsen.dk/hwlab/internal/ai"
	"git.georgsen.dk/hwlab/internal/inventory"
	"git.georgsen.dk/hwlab/internal/netbox"
)

// nonSafeChars matches characters that are not safe to send to SearXNG.
// Allowed: alphanumeric, space, dot, dash, underscore.
var nonSafeChars = regexp.MustCompile(`[^a-zA-Z0-9 .\-_]+`)

// SanitizeQuery strips unsafe characters from a search query string.
// Exported so it can be tested from the _test package.
func SanitizeQuery(s string) string {
	sanitized := nonSafeChars.ReplaceAllString(s, " ")
	return strings.TrimSpace(sanitized)
}

// NetBoxer is the subset of netbox.Client used by the Agent.
// Using an interface allows stub injection in tests.
type NetBoxer interface {
	ListDevicesWithStatus(ctx context.Context, status string) ([]netbox.Device, error)
	PatchCustomFields(ctx context.Context, deviceID int64, patch map[string]interface{}) error
}

// TextCompleter is the subset of ai.TierClient used by the Agent for text-only LLM calls.
type TextCompleter interface {
	TextComplete(ctx context.Context, prompt string) (string, error)
}

// CatalogTransitioner is the subset of inventory.CatalogUpdater used by the Agent.
type CatalogTransitioner interface {
	UpdateCatalogStatus(ctx context.Context, deviceID int64, current, next inventory.CatalogStatus) (inventory.CatalogStatus, error)
}

// Agent is the background worker that enriches needs_research hardware items.
// It polls NetBox, searches SearXNG, calls a Tier 2 LLM, and transitions items to researched.
type Agent struct {
	nbClient       NetBoxer
	researchClient ai.ResearchClient
	llm            TextCompleter
	updater        CatalogTransitioner
}

// NewAgent creates an Agent. All arguments must be non-nil.
func NewAgent(nb NetBoxer, rc ai.ResearchClient, llm TextCompleter, updater CatalogTransitioner) *Agent {
	return &Agent{
		nbClient:       nb,
		researchClient: rc,
		llm:            llm,
		updater:        updater,
	}
}

// enrichmentResponse is the expected JSON structure from the Tier 2 LLM.
type enrichmentResponse struct {
	AINotes    string `json:"ai_notes"`
	ProductURL string `json:"product_url"`
}

// RunOnce performs a single research cycle: finds all needs_research devices,
// enriches each via SearXNG + LLM, patches NetBox custom fields, and transitions
// the catalog status to researched. Returns the number of items enriched.
func (a *Agent) RunOnce(ctx context.Context) (int, error) {
	devices, err := a.nbClient.ListDevicesWithStatus(ctx, string(inventory.StatusNeedsResearch))
	if err != nil {
		return 0, fmt.Errorf("research agent: list needs_research devices: %w", err)
	}

	enriched := 0
	for _, dev := range devices {
		query := SanitizeQuery(dev.Name)
		if query == "" {
			log.Printf("research agent: device %d has empty name after sanitization, skipping", dev.ID)
			continue
		}

		results, err := a.researchClient.Search(ctx, query)
		if err != nil {
			log.Printf("research agent: search error for device %d (%q): %v", dev.ID, dev.Name, err)
			continue
		}
		if len(results) == 0 {
			log.Printf("research agent: no SearXNG results for device %d (%q), skipping", dev.ID, dev.Name)
			continue
		}

		// Build enrichment prompt using top 3 results
		top := results
		if len(top) > 3 {
			top = top[:3]
		}
		var sb strings.Builder
		for i, r := range top {
			sb.WriteString(fmt.Sprintf("%d. %s\n   %s\n   URL: %s\n", i+1, r.Title, r.Snippet, r.URL))
		}
		prompt := fmt.Sprintf(
			"You are enriching a hardware inventory record.\nItem: %s\nSearch results:\n%s\nReturn JSON: {\"ai_notes\": \"...\", \"product_url\": \"...\"}",
			dev.Name, sb.String(),
		)

		rawResponse, err := a.llm.TextComplete(ctx, prompt)
		if err != nil {
			log.Printf("research agent: LLM error for device %d (%q): %v", dev.ID, dev.Name, err)
			continue
		}

		// Parse LLM JSON response — extract ai_notes and product_url
		var enrichResp enrichmentResponse
		if parseErr := json.Unmarshal([]byte(rawResponse), &enrichResp); parseErr != nil {
			log.Printf("research agent: LLM non-JSON for device %d: %v (raw: %.100s)", dev.ID, parseErr, rawResponse)
			// Use raw response as ai_notes fallback
			enrichResp.AINotes = rawResponse
		}

		// Patch NetBox custom fields with enrichment data
		patch := map[string]interface{}{}
		if enrichResp.AINotes != "" {
			patch["ai_notes"] = enrichResp.AINotes
		}
		if enrichResp.ProductURL == "" && len(results) > 0 {
			// Fall back to first SearXNG result URL if LLM didn't provide one
			enrichResp.ProductURL = results[0].URL
		}
		if enrichResp.ProductURL != "" {
			patch["product_url"] = enrichResp.ProductURL
		}
		if len(patch) > 0 {
			if patchErr := a.nbClient.PatchCustomFields(ctx, int64(dev.ID), patch); patchErr != nil {
				log.Printf("research agent: patch error for device %d: %v", dev.ID, patchErr)
				continue
			}
		}

		// Transition catalog status: needs_research -> researched
		if _, transErr := a.updater.UpdateCatalogStatus(ctx, int64(dev.ID),
			inventory.StatusNeedsResearch, inventory.StatusResearched); transErr != nil {
			log.Printf("research agent: status transition error for device %d: %v", dev.ID, transErr)
			continue
		}

		enriched++
	}

	return enriched, nil
}

// Start runs the research agent on the given interval until ctx is cancelled.
// RunOnce is called immediately on start, then on each tick.
func (a *Agent) Start(ctx context.Context, interval time.Duration) {
	log.Printf("research agent: starting, interval=%v", interval)

	// Run immediately on startup
	if n, err := a.RunOnce(ctx); err != nil {
		log.Printf("research agent: initial cycle error: %v", err)
	} else {
		log.Printf("research agent: cycle complete, enriched %d items", n)
	}

	ticker := time.NewTicker(interval)
	defer ticker.Stop()

	for {
		select {
		case <-ctx.Done():
			log.Printf("research agent: shutting down")
			return
		case <-ticker.C:
			if n, err := a.RunOnce(ctx); err != nil {
				log.Printf("research agent: cycle error: %v", err)
			} else {
				log.Printf("research agent: cycle complete, enriched %d items", n)
			}
		}
	}
}