vxcontrol/pentagi · error · RetryableError

internal engine: summarization failed: %w

Error message

internal engine: summarization failed: %w

What it means

After pages are fetched, the internal engine builds a prompt (`buildPrompt`) and calls `e.summarizer(ctx, prompt)` — an LLM call that synthesizes the answer. If the summarizer returns an error, it is wrapped as Retryable with this message (backend/pkg/tools/searchers/internal.go:178), since an LLM outage is usually transient and the orchestrator can retry or fall back.

Source

Thrown at backend/pkg/tools/searchers/internal.go:178

		// [Source N]; usedURLs tracks the matching URL so it can be listed at the end.
		blocks = append(blocks, fmt.Sprintf("<source id=\"%d\" url=\"%s\">\n%s\n</source>", fetched+1, u, md))
		usedURLs = append(usedURLs, u)
		fetched++
	}

	if fetched == 0 {
		// Nothing could be read. If the pages errored (vs. returned empty), surface a
		// retryable error so the orchestrator can fall back to another engine.
		if lastFetch != nil {
			return "", Retryable(fmt.Errorf("internal engine: all pages failed to fetch: %w", lastFetch), 0)
		}
		return "", Fatal(fmt.Errorf("internal engine: no readable content found"))
	}

	prompt := e.buildPrompt(req.Query, blocks)
	answer, serr := e.summarizer(ctx, prompt)
	if serr != nil {
		return "", Retryable(fmt.Errorf("internal engine: summarization failed: %w", serr), 0)
	}

	return appendSources(answer, usedURLs), nil
}

// appendSources appends the list of source URLs the answer was synthesized from, so the
// [Source N] citations in the summary can be traced back to the page they came from. The
// numbering matches the source ids fed to the summarizer (1-based, in fetch order).
func appendSources(answer string, urls []string) string {
	if len(urls) == 0 {
		return answer
	}
	var sb strings.Builder
	sb.WriteString(strings.TrimRight(answer, "\n"))
	sb.WriteString("\n\n### Sources\n\n")
	for i, u := range urls {
		parsedURL, err := url.Parse(u)
		if err != nil {

View on GitHub (pinned to ea665308ba)

Solutions

1. Check the wrapped cause in the error message — it carries the provider's own error (401/429/5xx/timeout) and fix that specifically
2. Verify at least one valid LLM provider key is configured in .env / Settings UI
3. Retry — the error is typed Retryable, so transient provider failures are recoverable
4. Lower WebSearchInternalMaxSiteBytes (or WebSearchInternalMaxSites) to shrink the prompt if it is exceeding the model's context window

Example fix

// before: prompt too large for the model
WEB_SEARCH_INTERNAL_MAX_SITE_BYTES=1000000
// after
WEB_SEARCH_INTERNAL_MAX_SITE_BYTES=100000

// ensure a summarizer-capable provider is configured before invoking
if cfg.PrimaryProviderAPIKey == "" {
    return errors.New("no LLM provider configured for internal search summarizer")
}

_, err := engine.Handle(ctx, req)
if err != nil {
    if searchers.IsRetryable(err) {
        // backoff and retry; check wrapped provider error for 401/429/5xx
    }
}

When it happens

Trigger: The LLM backing the internal engine's summarizer fails: provider API key invalid/quota exhausted, model unavailable, request context deadline exceeded, rate limit (429), or the provider returns 5xx.

Common situations: No LLM provider key configured (or only optional ones set); provider quota/billing exhausted mid-run; LLM provider outage; context window exceeded because the fetched pages were large (WebSearchInternalMaxSiteBytes set too high).

Related errors

• failed to generate subtasks for task %d: %w
• summarization failed: %w
• failed to test provider: %w
• failed to call llm after %d retries: %w
• failed to get active task input: %w

AI-assisted analysis of vxcontrol/pentagi@ea665308ba (2026-09-01). Data as JSON: /api/errors/cdd0fc975adaaaa5. Report an issue: GitHub.