feat(summarizer): resilient endpoint chain with local→cloud fallback (ADR-022)
The first friendly-pilot live run produced zero summaries: koala/phi4-mini hit three silent failure modes — 8k context overflow on long transcripts (HTTP 400), intermittent malformed JSON (highlights as a bare string), and no fallback wired at all (summarizer.New(primary, nil)). Keep phi4-mini as the fast primary and add resilience around it: - Ordered endpoint chain (summarizer.NewChain): phi4-mini → koala/phi4-14b (local) → berget/mistral-small (worst-case external). All reached through the one LiteLLM gateway by alias. - A parse failure now advances the chain like a transport error — the old Primary→Fallback shape returned the parse error without trying anyone else. - Tolerant parse: highlights/takeaways coerce string→[]string, absorbing the common small-model quirk without spending a fallback round-trip. - Transcript truncation (TAPIR_MAX_TRANSCRIPT_CHARS=18000) prevents the overflow rather than recovering from it; validated to fit phi4-mini's 8k window. - Bounded completion budget (TAPIR_SUMMARY_MAX_TOKENS=1500) — the old 8192 budget itself contributed to the overflow. Local-first guarantee preserved by ordering: external endpoint is tried only after every local one fails. TAPIR_CLOUD_FALLBACK_MODEL="" disables it entirely for client/NDA deployments. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
@@ -1,17 +1,21 @@
|
||||
// Package summarizer implements ports.Summarizer backed by the copied llm
|
||||
// package's local-Primary -> BYO-Fallback routing (ADR-004). It is the only
|
||||
// place content ever leaves the engine toward an AI model, so it is also the
|
||||
// enforcement point for the local-first guarantee in
|
||||
// docs/use-cases/ai_routing.feature: a user with no BYO provider configured has
|
||||
// their content sent to the local stack and nowhere else.
|
||||
// package's routing (ADR-004, extended by ADR-022). It is the only place content
|
||||
// ever leaves the engine toward an AI model, so it is also the enforcement point
|
||||
// for the local-first guarantee in docs/use-cases/ai_routing.feature: endpoints
|
||||
// are tried in order, locals first, so content only reaches an external model
|
||||
// after every local endpoint has failed — and never at all when no external
|
||||
// endpoint is configured.
|
||||
package summarizer
|
||||
|
||||
import (
|
||||
"bytes"
|
||||
"context"
|
||||
"encoding/json"
|
||||
"errors"
|
||||
"fmt"
|
||||
"strings"
|
||||
"time"
|
||||
"unicode/utf8"
|
||||
|
||||
"gitea.d-ma.be/mathias/tapir/internal/domain"
|
||||
)
|
||||
@@ -29,19 +33,41 @@ type Endpoint struct {
|
||||
Model string // resolved alias, e.g. "iguana/deepseek-r1-14b"
|
||||
}
|
||||
|
||||
// Summarizer routes a transcript through the local endpoint first, then the
|
||||
// optional BYO endpoint. It owns its routing (rather than delegating to
|
||||
// llm.Router) so it can record which provider answered and whether the fallback
|
||||
// was used — information llm.Router collapses away.
|
||||
// Summarizer routes a transcript through an ordered chain of endpoints, trying
|
||||
// each in turn until one returns a parseable summary. It owns its routing
|
||||
// (rather than delegating to llm.Router) so it can record which provider answered
|
||||
// and whether a fallback was used — information llm.Router collapses away. The
|
||||
// chain ordering is the local-first guarantee: callers place local endpoints
|
||||
// first and any external endpoint last, so content only reaches an external model
|
||||
// after every local endpoint has failed.
|
||||
type Summarizer struct {
|
||||
primary Endpoint
|
||||
fallback *Endpoint // nil => no BYO; primary errors are returned, never sent externally
|
||||
now func() time.Time
|
||||
endpoints []Endpoint
|
||||
maxInputChars int // transcript truncation budget; 0 = no limit
|
||||
now func() time.Time
|
||||
}
|
||||
|
||||
// New constructs a Summarizer. fallback may be nil (no BYO provider configured).
|
||||
// New constructs a Summarizer from a primary endpoint and an optional fallback
|
||||
// (the historical local-Primary -> BYO-Fallback shape, ADR-004). A nil fallback
|
||||
// means a single-endpoint chain: errors are returned, content never leaves it.
|
||||
func New(primary Endpoint, fallback *Endpoint) *Summarizer {
|
||||
return &Summarizer{primary: primary, fallback: fallback, now: time.Now}
|
||||
eps := []Endpoint{primary}
|
||||
if fallback != nil {
|
||||
eps = append(eps, *fallback)
|
||||
}
|
||||
return &Summarizer{endpoints: eps, now: time.Now}
|
||||
}
|
||||
|
||||
// NewChain constructs a Summarizer over an ordered endpoint chain (ADR-022).
|
||||
// endpoints are tried in order; the first to return a parseable summary wins, and
|
||||
// FallbackUsed is recorded true for any endpoint past the first. maxInputChars
|
||||
// bounds the transcript text sent to every endpoint (0 = unbounded), so a long
|
||||
// transcript does not overflow a small-context primary model's window. It panics
|
||||
// on an empty chain — a wiring bug, not a runtime condition.
|
||||
func NewChain(endpoints []Endpoint, maxInputChars int) *Summarizer {
|
||||
if len(endpoints) == 0 {
|
||||
panic("summarizer: NewChain requires at least one endpoint")
|
||||
}
|
||||
return &Summarizer{endpoints: endpoints, maxInputChars: maxInputChars, now: time.Now}
|
||||
}
|
||||
|
||||
const systemPrompt = `You are Tapir, a video-summarization assistant.
|
||||
@@ -53,31 +79,35 @@ Respond with ONLY a JSON object, no prose and no code fences:
|
||||
- "takeaways": the actionable conclusions a viewer should leave with.
|
||||
Output the JSON object and nothing else.`
|
||||
|
||||
// Summarize implements ports.Summarizer.
|
||||
// Summarize implements ports.Summarizer. It walks the endpoint chain in order:
|
||||
// the first endpoint whose reply parses into a non-empty summary wins. An
|
||||
// endpoint is considered failed — and the next one tried — when the model call
|
||||
// errors OR when its reply cannot be parsed (a 200 with malformed JSON or a
|
||||
// highlights field the model emitted as a bare string). Truncation is applied
|
||||
// once, up front, so every endpoint sees the same bounded prompt. When the whole
|
||||
// chain fails, the joined error is returned so the engine queues the work for
|
||||
// retry and delivers no summary.
|
||||
func (s *Summarizer) Summarize(ctx context.Context, v domain.Video, t domain.Transcript) (domain.Summary, error) {
|
||||
if !t.HasText() {
|
||||
return domain.Summary{}, fmt.Errorf("summarize: transcript for video %s has no text", v.ID)
|
||||
}
|
||||
user := buildUserPrompt(v, t)
|
||||
user := buildUserPrompt(v, t, s.maxInputChars)
|
||||
|
||||
// Primary = local stack. Only on its failure is anything sent externally,
|
||||
// and only when a BYO fallback is configured.
|
||||
out, err := s.primary.Client.Complete(ctx, systemPrompt, user)
|
||||
if err == nil {
|
||||
return s.build(v, s.primary, false, out)
|
||||
var errs []error
|
||||
for i, ep := range s.endpoints {
|
||||
out, err := ep.Client.Complete(ctx, systemPrompt, user)
|
||||
if err != nil {
|
||||
errs = append(errs, fmt.Errorf("%s/%s call: %w", ep.Provider, ep.Model, err))
|
||||
continue
|
||||
}
|
||||
sum, perr := s.build(v, ep, i > 0, out)
|
||||
if perr != nil {
|
||||
errs = append(errs, fmt.Errorf("%s/%s output: %w", ep.Provider, ep.Model, perr))
|
||||
continue
|
||||
}
|
||||
return sum, nil
|
||||
}
|
||||
|
||||
if s.fallback == nil {
|
||||
// No BYO: content was sent to the local stack only. Surface the error so
|
||||
// the engine can queue the work for retry; deliver no summary.
|
||||
return domain.Summary{}, fmt.Errorf("summarize: local AI failed and no BYO provider configured: %w", err)
|
||||
}
|
||||
|
||||
out, ferr := s.fallback.Client.Complete(ctx, systemPrompt, user)
|
||||
if ferr != nil {
|
||||
return domain.Summary{}, fmt.Errorf("summarize: local AI failed: %w; BYO %s failed: %v", err, s.fallback.Provider, ferr)
|
||||
}
|
||||
return s.build(v, *s.fallback, true, out)
|
||||
return domain.Summary{}, fmt.Errorf("summarize: all %d endpoint(s) failed: %w", len(s.endpoints), errors.Join(errs...))
|
||||
}
|
||||
|
||||
func (s *Summarizer) build(v domain.Video, ep Endpoint, fallbackUsed bool, raw string) (domain.Summary, error) {
|
||||
@@ -89,8 +119,8 @@ func (s *Summarizer) build(v domain.Video, ep Endpoint, fallbackUsed bool, raw s
|
||||
UserID: v.UserID,
|
||||
VideoID: v.ID,
|
||||
Summary: parsed.Summary,
|
||||
Highlights: parsed.Highlights,
|
||||
Takeaways: parsed.Takeaways,
|
||||
Highlights: []string(parsed.Highlights),
|
||||
Takeaways: []string(parsed.Takeaways),
|
||||
AIProvider: ep.Provider,
|
||||
AIModel: ep.Model,
|
||||
FallbackUsed: fallbackUsed,
|
||||
@@ -98,20 +128,94 @@ func (s *Summarizer) build(v domain.Video, ep Endpoint, fallbackUsed bool, raw s
|
||||
}, nil
|
||||
}
|
||||
|
||||
func buildUserPrompt(v domain.Video, t domain.Transcript) string {
|
||||
func buildUserPrompt(v domain.Video, t domain.Transcript, maxInputChars int) string {
|
||||
var b strings.Builder
|
||||
fmt.Fprintf(&b, "Title: %s\n", v.Title)
|
||||
if v.URL != "" {
|
||||
fmt.Fprintf(&b, "URL: %s\n", v.URL)
|
||||
}
|
||||
fmt.Fprintf(&b, "\nTranscript:\n%s", t.Content)
|
||||
fmt.Fprintf(&b, "\nTranscript:\n%s", truncate(t.Content, maxInputChars))
|
||||
return b.String()
|
||||
}
|
||||
|
||||
// truncate caps content to max bytes on a UTF-8 rune boundary, appending a
|
||||
// marker so the model knows the transcript was cut. A non-positive max (or a
|
||||
// content already within budget) returns content unchanged. Bounding the input
|
||||
// keeps a long transcript from overflowing a small-context model's window — the
|
||||
// production failure mode where koala/phi4-mini's 8k context returned HTTP 400 on
|
||||
// a 11.6k-token transcript.
|
||||
func truncate(content string, max int) string {
|
||||
if max <= 0 || len(content) <= max {
|
||||
return content
|
||||
}
|
||||
cut := max
|
||||
for cut > 0 && !utf8.RuneStart(content[cut]) {
|
||||
cut--
|
||||
}
|
||||
return content[:cut] + "\n…[transcript truncated to fit the model context]"
|
||||
}
|
||||
|
||||
// flexStrings is a []string that also unmarshals from a single JSON string or a
|
||||
// JSON array of scalars. Small local models (koala/phi4-mini) sometimes emit
|
||||
// "highlights": "one point" instead of an array, or mix in a number; rather than
|
||||
// fail the whole summary on that quirk, coerce to []string. Empty/whitespace
|
||||
// elements are dropped.
|
||||
type flexStrings []string
|
||||
|
||||
func (f *flexStrings) UnmarshalJSON(b []byte) error {
|
||||
b = bytes.TrimSpace(b)
|
||||
if len(b) == 0 || string(b) == "null" {
|
||||
*f = nil
|
||||
return nil
|
||||
}
|
||||
if b[0] == '[' {
|
||||
var raw []json.RawMessage
|
||||
if err := json.Unmarshal(b, &raw); err != nil {
|
||||
return err
|
||||
}
|
||||
out := make([]string, 0, len(raw))
|
||||
for _, r := range raw {
|
||||
s, err := rawToString(r)
|
||||
if err != nil {
|
||||
return err
|
||||
}
|
||||
if strings.TrimSpace(s) != "" {
|
||||
out = append(out, s)
|
||||
}
|
||||
}
|
||||
*f = out
|
||||
return nil
|
||||
}
|
||||
s, err := rawToString(b)
|
||||
if err != nil {
|
||||
return err
|
||||
}
|
||||
if strings.TrimSpace(s) == "" {
|
||||
*f = nil
|
||||
} else {
|
||||
*f = flexStrings{s}
|
||||
}
|
||||
return nil
|
||||
}
|
||||
|
||||
// rawToString renders a JSON scalar as text: a quoted string is unquoted; any
|
||||
// other scalar (number, bool) is kept as its literal source so no content is lost.
|
||||
func rawToString(r json.RawMessage) (string, error) {
|
||||
r = bytes.TrimSpace(r)
|
||||
if len(r) > 0 && r[0] == '"' {
|
||||
var s string
|
||||
if err := json.Unmarshal(r, &s); err != nil {
|
||||
return "", err
|
||||
}
|
||||
return s, nil
|
||||
}
|
||||
return string(r), nil
|
||||
}
|
||||
|
||||
type parsedSummary struct {
|
||||
Summary string `json:"summary"`
|
||||
Highlights []string `json:"highlights"`
|
||||
Takeaways []string `json:"takeaways"`
|
||||
Summary string `json:"summary"`
|
||||
Highlights flexStrings `json:"highlights"`
|
||||
Takeaways flexStrings `json:"takeaways"`
|
||||
}
|
||||
|
||||
// parse extracts the JSON object from a model reply. Thinking models (qwen3,
|
||||
|
||||
@@ -129,8 +129,8 @@ func TestSummarize_NoBYO_ContentOnlyLocal(t *testing.T) {
|
||||
local := &fakeClient{reply: goodReply}
|
||||
s := New(Endpoint{Client: local, Provider: "local", Model: "iguana/deepseek-r1-14b"}, nil)
|
||||
|
||||
if s.fallback != nil {
|
||||
t.Fatal("no BYO configured but fallback endpoint is non-nil")
|
||||
if len(s.endpoints) != 1 {
|
||||
t.Fatalf("no BYO configured but chain has %d endpoints, want 1", len(s.endpoints))
|
||||
}
|
||||
for i := 0; i < 3; i++ {
|
||||
sum, err := s.Summarize(context.Background(), testVideo(), testTranscript())
|
||||
@@ -176,3 +176,97 @@ func TestParse_EmptySummaryRejected(t *testing.T) {
|
||||
t.Fatal("want error for empty summary (thinking model returned no content)")
|
||||
}
|
||||
}
|
||||
|
||||
// parse tolerates a small model emitting "highlights" as a bare string instead
|
||||
// of an array — the production koala/phi4-mini quirk that errored with
|
||||
// "cannot unmarshal string into Go struct field ... highlights of type []string".
|
||||
func TestParse_ToleratesStringHighlights(t *testing.T) {
|
||||
p, err := parse(`{"summary":"s","highlights":"one big point","takeaways":["a","b"]}`)
|
||||
if err != nil {
|
||||
t.Fatalf("parse: %v", err)
|
||||
}
|
||||
if len(p.Highlights) != 1 || p.Highlights[0] != "one big point" {
|
||||
t.Errorf("highlights = %v, want [\"one big point\"]", p.Highlights)
|
||||
}
|
||||
if len(p.Takeaways) != 2 {
|
||||
t.Errorf("takeaways = %v, want 2", p.Takeaways)
|
||||
}
|
||||
}
|
||||
|
||||
// Chain: an endpoint that returns a 200 with unparseable output is treated as a
|
||||
// failure, and the next endpoint in the chain is tried. This is the case the old
|
||||
// primary->fallback shape missed — a parse error short-circuited instead of
|
||||
// falling back.
|
||||
func TestSummarize_FallsBackOnMalformedOutput(t *testing.T) {
|
||||
bad := &fakeClient{reply: `{"summary": not json`}
|
||||
good := &fakeClient{reply: goodReply}
|
||||
s := NewChain([]Endpoint{
|
||||
{Client: bad, Provider: "local", Model: "koala/phi4-mini"},
|
||||
{Client: good, Provider: "local", Model: "koala/phi4-14b"},
|
||||
}, 0)
|
||||
|
||||
sum, err := s.Summarize(context.Background(), testVideo(), testTranscript())
|
||||
if err != nil {
|
||||
t.Fatalf("Summarize: %v", err)
|
||||
}
|
||||
if sum.AIModel != "koala/phi4-14b" {
|
||||
t.Errorf("AIModel = %q, want koala/phi4-14b (fell back past malformed primary)", sum.AIModel)
|
||||
}
|
||||
if !sum.FallbackUsed {
|
||||
t.Error("FallbackUsed = false, want true")
|
||||
}
|
||||
if bad.calls != 1 || good.calls != 1 {
|
||||
t.Errorf("calls: bad=%d good=%d, want 1 and 1", bad.calls, good.calls)
|
||||
}
|
||||
}
|
||||
|
||||
// Chain: when every endpoint fails, no summary is produced and the joined error
|
||||
// names each failure so the engine queues the work for retry.
|
||||
func TestSummarize_ChainAllEndpointsFail(t *testing.T) {
|
||||
a := &fakeClient{err: errors.New("context overflow")}
|
||||
b := &fakeClient{reply: "not even json"}
|
||||
s := NewChain([]Endpoint{
|
||||
{Client: a, Provider: "local", Model: "m1"},
|
||||
{Client: b, Provider: "berget", Model: "m2"},
|
||||
}, 0)
|
||||
|
||||
if _, err := s.Summarize(context.Background(), testVideo(), testTranscript()); err == nil {
|
||||
t.Fatal("want error when all endpoints fail")
|
||||
}
|
||||
if a.calls != 1 || b.calls != 1 {
|
||||
t.Errorf("calls: a=%d b=%d, want 1 and 1", a.calls, b.calls)
|
||||
}
|
||||
}
|
||||
|
||||
// A transcript longer than the chain's input budget is truncated before it
|
||||
// reaches any model, so a small-context primary does not overflow its window.
|
||||
func TestSummarize_TruncatesLongTranscript(t *testing.T) {
|
||||
local := &fakeClient{reply: goodReply}
|
||||
const budget = 100
|
||||
s := NewChain([]Endpoint{{Client: local, Provider: "local", Model: "m"}}, budget)
|
||||
|
||||
long := domain.Transcript{
|
||||
VideoID: "vid-1", UserID: "user-1", Source: domain.SourceCaptions,
|
||||
Content: strings.Repeat("word ", 1000), // 5000 bytes, well over budget
|
||||
}
|
||||
if _, err := s.Summarize(context.Background(), testVideo(), long); err != nil {
|
||||
t.Fatalf("Summarize: %v", err)
|
||||
}
|
||||
// The prompt carries title/URL framing plus the truncation marker, so allow
|
||||
// headroom over the raw transcript budget — but it must be far below 5000.
|
||||
if len(local.lastUser) > budget+300 {
|
||||
t.Errorf("prompt length = %d, want <= %d (transcript not truncated)", len(local.lastUser), budget+300)
|
||||
}
|
||||
if !strings.Contains(local.lastUser, "truncated") {
|
||||
t.Error("truncation marker missing from prompt")
|
||||
}
|
||||
}
|
||||
|
||||
func TestNewChain_PanicsOnEmptyChain(t *testing.T) {
|
||||
defer func() {
|
||||
if recover() == nil {
|
||||
t.Fatal("want panic on empty endpoint chain")
|
||||
}
|
||||
}()
|
||||
NewChain(nil, 0)
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user