Add semantic-cache module: CallAdvisor over a Redis Stack vector index with a real in-process embedding model; threshold sweep (hit rate against wrong answers), 600-request replay with an illustrative cost model, tenant filter, expiry, look-alike and embedding-model-change traps; optional real local model
Co-Authored-By: Claude Sonnet 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_01G8ikz8xdWuTP5yun8DZ1hk
This commit is contained in:
Claude Sonnet 5.5
@@ -0,0 +1,580 @@
|
||||
package com.ankurm.semanticcache;
|
||||
|
||||
import static org.assertj.core.api.Assertions.assertThat;
|
||||
|
||||
import java.util.ArrayList;
|
||||
import java.util.Comparator;
|
||||
import java.util.List;
|
||||
import java.util.Map;
|
||||
import java.util.Random;
|
||||
|
||||
import org.junit.jupiter.api.BeforeAll;
|
||||
import org.junit.jupiter.api.Test;
|
||||
import org.springframework.ai.chat.client.ChatClient;
|
||||
import org.springframework.ai.embedding.EmbeddingModel;
|
||||
import redis.clients.jedis.RedisClient;
|
||||
|
||||
/**
|
||||
* Every test writes one transcript under output/ and asserts the same numbers it prints. Needs a Redis with the
|
||||
* RediSearch module on 127.0.0.1:6393 (scripts/services-up.sh starts one).
|
||||
*/
|
||||
class SemanticCacheTest {
|
||||
|
||||
static final int REDIS_PORT = 6393;
|
||||
static final String TENANT = "shop-a";
|
||||
|
||||
static EmbeddingModel embeddings;
|
||||
static RedisClient jedis;
|
||||
static Dataset data;
|
||||
static Map<String, String> intentByText;
|
||||
|
||||
@BeforeAll
|
||||
static void setUp() {
|
||||
embeddings = Embedder.miniLm();
|
||||
jedis = RedisClient.create("127.0.0.1", REDIS_PORT);
|
||||
data = new Dataset();
|
||||
intentByText = data.intentByText();
|
||||
}
|
||||
|
||||
static SemanticCache freshCache(double threshold) {
|
||||
jedis.flushAll();
|
||||
return new SemanticCache(jedis, embeddings, threshold);
|
||||
}
|
||||
|
||||
static SemanticCache seededCache(double threshold) {
|
||||
SemanticCache c = freshCache(threshold);
|
||||
data.seeds().forEach(r -> c.put(TENANT, r.text(), ScriptedModel.answerFor(r.intent())));
|
||||
return c;
|
||||
}
|
||||
|
||||
static String intentOfAnswer(String answer) {
|
||||
return answer.substring(1, answer.indexOf(']'));
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------- 1. what a similarity score looks like
|
||||
|
||||
@Test
|
||||
void similarityScoresOfParaphrasesAndLookAlikes() {
|
||||
List<Double> paraphrase = new ArrayList<>(), lookAlike = new ArrayList<>(), offTopic = new ArrayList<>();
|
||||
Map<String, float[]> seedVec = new java.util.LinkedHashMap<>();
|
||||
data.seeds().forEach(r -> seedVec.put(r.intent(), embeddings.embed(r.text())));
|
||||
String worstLookAlike = "";
|
||||
double worstScore = -1;
|
||||
String lowestParaphrase = "";
|
||||
double lowestScore = 2;
|
||||
for (var r : data.paraphrases()) {
|
||||
float[] v = embeddings.embed(r.text());
|
||||
double own = Embedder.cosine(v, seedVec.get(r.intent()));
|
||||
paraphrase.add(own);
|
||||
if (own < lowestScore) {
|
||||
lowestScore = own;
|
||||
lowestParaphrase = r.text() + " -> " + seedText(r.intent());
|
||||
}
|
||||
for (var e : seedVec.entrySet()) {
|
||||
if (e.getKey().equals(r.intent())) continue;
|
||||
double s = Embedder.cosine(v, e.getValue());
|
||||
if (s > worstScore) {
|
||||
worstScore = s;
|
||||
worstLookAlike = r.text() + " -> " + seedText(e.getKey());
|
||||
}
|
||||
}
|
||||
double best = seedVec.entrySet().stream().filter(e -> !e.getKey().equals(r.intent()))
|
||||
.mapToDouble(e -> Embedder.cosine(v, e.getValue())).max().orElse(0);
|
||||
lookAlike.add(best);
|
||||
}
|
||||
for (var r : data.offTopic()) {
|
||||
float[] v = embeddings.embed(r.text());
|
||||
offTopic.add(seedVec.values().stream().mapToDouble(s -> Embedder.cosine(v, s)).max().orElse(0));
|
||||
}
|
||||
|
||||
// Which different intents are closest to each other? Every pair of cached seeds, 30 x 29 / 2 = 435 pairs.
|
||||
record Pair(String a, String b, double score) { }
|
||||
List<Pair> pairs = new ArrayList<>();
|
||||
List<String> seedIntents = new ArrayList<>(seedVec.keySet());
|
||||
for (int i = 0; i < seedIntents.size(); i++)
|
||||
for (int j = i + 1; j < seedIntents.size(); j++)
|
||||
pairs.add(new Pair(seedIntents.get(i), seedIntents.get(j),
|
||||
(1 + Embedder.cosine(seedVec.get(seedIntents.get(i)), seedVec.get(seedIntents.get(j)))) / 2));
|
||||
pairs.sort(Comparator.comparingDouble(Pair::score).reversed());
|
||||
|
||||
// Does Redis report the same number? Ask for one question and compare with the cosine computed in Java.
|
||||
SemanticCache cache = seededCache(0.0);
|
||||
var probe = data.paraphrases().get(0);
|
||||
var redisHit = cache.nearest(TENANT, probe.text()).orElseThrow();
|
||||
double javaCosine = Embedder.cosine(embeddings.embed(probe.text()), seedVec.get(probe.intent()));
|
||||
var exact = cache.nearest(TENANT, data.seeds().get(0).text()).orElseThrow();
|
||||
|
||||
try (var t = new Transcript("01-similarity-scores.txt", "What a similarity score looks like (all-MiniLM-L6-v2, cosine)")) {
|
||||
t.line("Plain cosine similarity (NOT the Spring AI score, see the end) of each later phrasing with the cached phrasing of the SAME question (%d pairs):", paraphrase.size());
|
||||
t.line(" %s", summary(paraphrase));
|
||||
t.line(" lowest: %.3f %s", lowestScore, lowestParaphrase);
|
||||
t.blank();
|
||||
t.line("Best score each of those phrasings gets against a cached question of a DIFFERENT intent (look-alikes like reset password / reset router):");
|
||||
t.line(" %s", summary(lookAlike));
|
||||
t.line(" highest: %.3f %s", worstScore, worstLookAlike);
|
||||
t.blank();
|
||||
t.line("Best score of %d off-topic questions against any cached question:", offTopic.size());
|
||||
t.line(" %s", summary(offTopic));
|
||||
t.blank();
|
||||
t.line("Does Redis report the same number? \"%s\"", probe.text());
|
||||
t.line(" cosine computed in Java : %.6f", javaCosine);
|
||||
t.line(" Document.getScore() from Redis : %.6f (matched stored question: \"%s\")", redisHit.score(), redisHit.storedQuestion());
|
||||
t.line(" (1 + cosine) / 2 : %.6f <- the score is this, not the cosine", (1 + javaCosine) / 2);
|
||||
t.line(" same question as the cached one -> score %.6f", exact.score());
|
||||
t.blank();
|
||||
t.line("The closest pairs of DIFFERENT cached questions (Spring AI score; 435 pairs in all). Above the threshold, one answers for the other:");
|
||||
for (int i = 0; i < 8; i++)
|
||||
t.line(" %.3f \"%s\" / \"%s\"", pairs.get(i).score(), seedText(pairs.get(i).a()), seedText(pairs.get(i).b()));
|
||||
t.blank();
|
||||
t.line("Spring AI's similarityThreshold is applied to that score. Threshold -> the cosine it really means (cosine = 2 x score - 1):");
|
||||
for (double th : new double[]{0.70, 0.80, 0.85, 0.90, 0.95}) t.line(" similarityThreshold %.2f = cosine %.2f", th, 2 * th - 1);
|
||||
}
|
||||
assertThat(redisHit.score()).isCloseTo((1 + javaCosine) / 2, org.assertj.core.data.Offset.offset(1e-3));
|
||||
assertThat(exact.score()).isGreaterThan(0.999);
|
||||
// the two populations overlap: the reason a threshold cannot be both safe and generous
|
||||
assertThat(max(lookAlike)).isGreaterThan(min(paraphrase));
|
||||
assertThat(pairs.get(0).score()).isGreaterThan(0.80);
|
||||
}
|
||||
|
||||
static String seedText(String intent) {
|
||||
return data.seeds().stream().filter(r -> r.intent().equals(intent)).findFirst().orElseThrow().text();
|
||||
}
|
||||
|
||||
static double min(List<Double> v) { return v.stream().mapToDouble(Double::doubleValue).min().orElse(0); }
|
||||
|
||||
static double max(List<Double> v) { return v.stream().mapToDouble(Double::doubleValue).max().orElse(0); }
|
||||
|
||||
static String summary(List<Double> v) {
|
||||
List<Double> s = v.stream().sorted().toList();
|
||||
return String.format("min %.3f p10 %.3f median %.3f p90 %.3f max %.3f", s.get(0), s.get((int) (s.size() * 0.1)),
|
||||
s.get(s.size() / 2), s.get((int) (s.size() * 0.9) - 1), s.get(s.size() - 1));
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------- 2. the advisor in a ChatClient
|
||||
|
||||
@Test
|
||||
void theAdvisorServesAParaphraseWithoutCallingTheModel() {
|
||||
SemanticCache cache = freshCache(0.80);
|
||||
ScriptedModel model = new ScriptedModel(intentByText, 0);
|
||||
SemanticCacheAdvisor advisor = new SemanticCacheAdvisor(cache, 0);
|
||||
ChatClient client = ChatClient.builder(model).defaultAdvisors(advisor).build();
|
||||
|
||||
String[][] asks = {
|
||||
{"reset-password", "How do I reset my account password?"},
|
||||
{"reset-password", "How do I reset my account password?"},
|
||||
{"reset-password", "I forgot my password, how can I get back into my account?"},
|
||||
{"reset-router", "How do I reset my router to factory settings?"},
|
||||
{"ood", "What is the capital of Australia?"},
|
||||
};
|
||||
try (var t = new Transcript("02-advisor-in-a-chat-client.txt", "SemanticCacheAdvisor in front of a ChatClient (threshold 0.80)")) {
|
||||
for (String[] a : asks) {
|
||||
long before = model.calls.get();
|
||||
var resp = client.prompt().user(a[1]).call().chatClientResponse();
|
||||
boolean hit = Boolean.TRUE.equals(resp.context().get(SemanticCacheAdvisor.HIT));
|
||||
String answer = resp.chatResponse().getResult().getOutput().getText();
|
||||
t.line("ask : %s", a[1]);
|
||||
t.line(" -> %s%s, model calls so far %d, answer starts \"%s\"", hit ? "CACHE HIT, score " : "model call",
|
||||
hit ? String.format("%.3f", (Double) resp.context().get(SemanticCacheAdvisor.SCORE)) : "", model.calls.get(),
|
||||
answer.substring(0, answer.indexOf(']') + 1));
|
||||
assertThat(model.calls.get() - before).isEqualTo(hit ? 0 : 1);
|
||||
}
|
||||
t.blank();
|
||||
t.line("advisor counters: %d hits, %d misses; model received %d calls for %d questions; entries in Redis: %d",
|
||||
advisor.hits.get(), advisor.misses.get(), model.calls.get(), asks.length, cache.size());
|
||||
}
|
||||
assertThat(advisor.hits.get()).isEqualTo(2);
|
||||
assertThat(model.calls.get()).isEqualTo(3);
|
||||
assertThat(cache.size()).isEqualTo(3);
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------- 3. the threshold sweep
|
||||
|
||||
record Probe(String intent, String text, boolean offTopic, SemanticCache.Hit nearest) {
|
||||
String nearestIntent() { return nearest == null ? "-" : intentOfAnswer(nearest.answer()); }
|
||||
}
|
||||
|
||||
static List<Probe> probes(SemanticCache cache) {
|
||||
List<Probe> out = new ArrayList<>();
|
||||
for (var r : data.paraphrases()) out.add(new Probe(r.intent(), r.text(), false, cache.nearest(TENANT, r.text()).orElse(null)));
|
||||
for (var r : data.offTopic()) out.add(new Probe(r.intent(), r.text(), true, cache.nearest(TENANT, r.text()).orElse(null)));
|
||||
return out;
|
||||
}
|
||||
|
||||
@Test
|
||||
void thresholdSweepHitRateAgainstWrongAnswers() {
|
||||
SemanticCache cache = seededCache(0.0);
|
||||
List<Probe> probes = probes(cache);
|
||||
long nPara = probes.stream().filter(p -> !p.offTopic()).count();
|
||||
long nOff = probes.size() - nPara;
|
||||
double[] thresholds = {0.50, 0.60, 0.70, 0.75, 0.80, 0.85, 0.90, 0.95};
|
||||
List<float[]> sv = data.seeds().stream().map(r -> embeddings.embed(r.text())).toList();
|
||||
int[] confusable = new int[thresholds.length];
|
||||
for (int i = 0; i < sv.size(); i++)
|
||||
for (int j = i + 1; j < sv.size(); j++) {
|
||||
double sc = (1 + Embedder.cosine(sv.get(i), sv.get(j))) / 2;
|
||||
for (int k = 0; k < thresholds.length; k++) if (sc >= thresholds[k]) confusable[k]++;
|
||||
}
|
||||
try (var t = new Transcript("03-threshold-sweep.txt", "Threshold sweep: 30 cached questions, 90 rewordings and 20 off-topic questions asked")) {
|
||||
t.line("threshold = the value given to SearchRequest.similarityThreshold (Spring AI score); cosine = 2 x threshold - 1.");
|
||||
t.line("A served answer is RIGHT if it belongs to the intent of the question asked, WRONG if it belongs to another intent.");
|
||||
t.line("An off-topic question has no right answer in the cache, so any hit on it is WRONG.");
|
||||
t.blank();
|
||||
t.line("%9s | %6s | %8s | %9s | %14s | %15s | %s | %s", "threshold", "cosine", "hit rate", "right", "wrong (reworded)", "wrong (off-topic)", "wrong / all hits", "cached pairs of different intents that answer for each other");
|
||||
for (int ti = 0; ti < thresholds.length; ti++) {
|
||||
double th = thresholds[ti];
|
||||
long hits = 0, right = 0, wrongPara = 0, wrongOff = 0, paraHits = 0;
|
||||
for (Probe p : probes) {
|
||||
if (p.nearest() == null || p.nearest().score() < th) continue;
|
||||
hits++;
|
||||
boolean ok = !p.offTopic() && p.nearestIntent().equals(p.intent());
|
||||
if (!p.offTopic()) paraHits++;
|
||||
if (ok) right++;
|
||||
else if (p.offTopic()) wrongOff++;
|
||||
else wrongPara++;
|
||||
}
|
||||
t.line("%9.2f | %6.2f | %7.0f%% | %5d/%-3d | %8d/%-6d | %9d/%-7d | %s | %s", th, 2 * th - 1, 100.0 * paraHits / nPara, right, nPara, wrongPara, nPara,
|
||||
wrongOff, nOff, hits == 0 ? "-" : String.format("%.1f%%", 100.0 * (wrongPara + wrongOff) / hits), confusable[ti] + " of 435");
|
||||
}
|
||||
t.blank();
|
||||
t.line("Questions that were answered WRONGLY at threshold 0.80 (asked -> served from): %s", probes.stream().anyMatch(p -> p.nearest().score() >= 0.80 && (p.offTopic() || !p.nearestIntent().equals(p.intent()))) ? "" : "none among these 110 probes");
|
||||
for (Probe p : probes) {
|
||||
if (p.nearest() != null && p.nearest().score() >= 0.80 && (p.offTopic() || !p.nearestIntent().equals(p.intent())))
|
||||
t.line(" %.3f \"%s\" -> \"%s\"", p.nearest().score(), p.text(), p.nearest().storedQuestion());
|
||||
}
|
||||
t.blank();
|
||||
t.line("Rewordings the cache MISSED at threshold 0.80 (best stored question and its score):");
|
||||
for (Probe p : probes) {
|
||||
if (!p.offTopic() && (p.nearest() == null || p.nearest().score() < 0.80))
|
||||
t.line(" %.3f \"%s\" (nearest: \"%s\")", p.nearest().score(), p.text(), p.nearest().storedQuestion());
|
||||
}
|
||||
}
|
||||
// the two ends of the dial
|
||||
long hitsAt50 = probes.stream().filter(p -> p.nearest().score() >= 0.50).count();
|
||||
long offHitsAt50 = probes.stream().filter(p -> p.offTopic() && p.nearest().score() >= 0.50).count();
|
||||
long wrongAt95 = probes.stream().filter(p -> p.nearest().score() >= 0.95 && (p.offTopic() || !p.nearestIntent().equals(p.intent()))).count();
|
||||
assertThat(hitsAt50).isGreaterThan(nPara);
|
||||
assertThat(offHitsAt50).isGreaterThan(0);
|
||||
assertThat(wrongAt95).isZero();
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------- 4. replay a workload
|
||||
|
||||
record Replay(double threshold, int requests, long modelCalls, long hits, long wrong, long promptTokens, long completionTokens,
|
||||
double avgLookupMillis) { }
|
||||
|
||||
static Replay replay(double threshold, List<Data> workload) {
|
||||
SemanticCache cache = freshCache(threshold);
|
||||
ScriptedModel model = new ScriptedModel(intentByText, 0);
|
||||
SemanticCacheAdvisor advisor = new SemanticCacheAdvisor(cache, 0);
|
||||
ChatClient client = ChatClient.builder(model).defaultAdvisors(advisor).build();
|
||||
long wrong = 0;
|
||||
long lookupNanos = 0;
|
||||
for (Data d : workload) {
|
||||
long t0 = System.nanoTime();
|
||||
var resp = client.prompt().user(d.text()).call().chatClientResponse();
|
||||
long elapsed = System.nanoTime() - t0;
|
||||
boolean hit = Boolean.TRUE.equals(resp.context().get(SemanticCacheAdvisor.HIT));
|
||||
if (hit) {
|
||||
lookupNanos += elapsed;
|
||||
String served = intentOfAnswer(resp.chatResponse().getResult().getOutput().getText());
|
||||
if (!served.equals(d.intent())) wrong++;
|
||||
}
|
||||
}
|
||||
double avgHitMillis = advisor.hits.get() == 0 ? 0 : lookupNanos / 1e6 / advisor.hits.get();
|
||||
return new Replay(threshold, workload.size(), model.calls.get(), advisor.hits.get(), wrong, model.promptTokens.get(),
|
||||
model.completionTokens.get(), avgHitMillis);
|
||||
}
|
||||
|
||||
record Data(String intent, String text) { }
|
||||
|
||||
/** 600 requests: popular intents asked far more often, each time in a random one of its four phrasings, 15% off-topic. */
|
||||
static List<Data> workload() {
|
||||
Random rnd = new Random(42);
|
||||
List<String> intents = data.seeds().stream().map(Dataset.Row::intent).toList();
|
||||
Map<String, List<String>> phrasings = new java.util.LinkedHashMap<>();
|
||||
data.rows.stream().filter(r -> !r.role().equals("ood")).forEach(r -> phrasings.computeIfAbsent(r.intent(), k -> new ArrayList<>()).add(r.text()));
|
||||
double[] weight = new double[intents.size()];
|
||||
double total = 0;
|
||||
for (int i = 0; i < weight.length; i++) {
|
||||
weight[i] = 1.0 / (i + 1); // Zipf: the first intent is asked 30 times as often as the last
|
||||
total += weight[i];
|
||||
}
|
||||
List<Data> out = new ArrayList<>();
|
||||
List<Dataset.Row> off = data.offTopic();
|
||||
for (int n = 0; n < 600; n++) {
|
||||
if (rnd.nextDouble() < 0.15) {
|
||||
var r = off.get(rnd.nextInt(off.size()));
|
||||
out.add(new Data(r.intent(), r.text()));
|
||||
continue;
|
||||
}
|
||||
double x = rnd.nextDouble() * total;
|
||||
int i = 0;
|
||||
while (i < weight.length - 1 && (x -= weight[i]) > 0) i++;
|
||||
List<String> ph = phrasings.get(intents.get(i));
|
||||
out.add(new Data(intents.get(i), ph.get(rnd.nextInt(ph.size()))));
|
||||
}
|
||||
return out;
|
||||
}
|
||||
|
||||
static final double PRICE_IN_PER_MILLION = 2.50, PRICE_OUT_PER_MILLION = 10.00; // illustrative, not a vendor's price sheet
|
||||
|
||||
static double dollars(long in, long out) {
|
||||
return in / 1e6 * PRICE_IN_PER_MILLION + out / 1e6 * PRICE_OUT_PER_MILLION;
|
||||
}
|
||||
|
||||
@Test
|
||||
void replayingSixHundredRequests() {
|
||||
List<Data> workload = workload();
|
||||
long distinctTexts = workload.stream().map(Data::text).distinct().count();
|
||||
long exactRepeats = workload.size() - distinctTexts;
|
||||
ScriptedModel baseline = new ScriptedModel(intentByText, 0);
|
||||
ChatClient plain = ChatClient.builder(baseline).build();
|
||||
workload.forEach(d -> plain.prompt().user(d.text()).call().content());
|
||||
double baseCost = dollars(baseline.promptTokens.get(), baseline.completionTokens.get());
|
||||
|
||||
List<Replay> runs = new ArrayList<>();
|
||||
for (double th : new double[]{0.70, 0.80, 0.90, 0.95, 0.9999}) runs.add(replay(th, workload));
|
||||
|
||||
try (var t = new Transcript("04-replay-600-requests.txt", "600 requests through the advisor, by threshold")) {
|
||||
t.line("Workload: 600 requests (seed 42), 30 intents asked with Zipf popularity in random phrasings, 15% off-topic.");
|
||||
t.line(" %d distinct texts, so a plain exact-match cache could answer at most %d of 600 (%.0f%%).", distinctTexts, exactRepeats, 100.0 * exactRepeats / 600);
|
||||
t.line("Cost model (ILLUSTRATIVE): $%.2f per million input tokens, $%.2f per million output tokens, tokens = characters / 4.", PRICE_IN_PER_MILLION, PRICE_OUT_PER_MILLION);
|
||||
t.blank();
|
||||
t.line("%-22s | %10s | %8s | %6s | %13s | %9s | %s", "setup", "model calls", "hit rate", "wrong", "wrong of hits", "model $", "saved");
|
||||
t.line("%-22s | %10d | %8s | %6s | %13s | %9.4f | %s", "no cache", baseline.calls.get(), "-", "-", "-", baseCost, "-");
|
||||
for (Replay r : runs) {
|
||||
double cost = dollars(r.promptTokens, r.completionTokens);
|
||||
t.line("%-22s | %10d | %7.1f%% | %6d | %12.1f%% | %9.4f | %.0f%%",
|
||||
r.threshold >= 0.999 ? "cache, exact match only" : String.format("cache, threshold %.2f", r.threshold), r.modelCalls,
|
||||
100.0 * r.hits / r.requests, r.wrong, r.hits == 0 ? 0 : 100.0 * r.wrong / r.hits, cost, 100.0 * (1 - cost / baseCost));
|
||||
}
|
||||
t.blank();
|
||||
t.line("Average time to serve a cache hit (embed the question, search Redis, build the response): %s",
|
||||
runs.stream().filter(r -> r.threshold < 0.999).map(r -> String.format("%.0f ms at %.2f", r.avgLookupMillis, r.threshold)).reduce((a, b) -> a + ", " + b).orElse(""));
|
||||
t.line("Saved dollars exclude the cost of embedding (local model here, CPU only) and of running Redis.");
|
||||
}
|
||||
Replay at80 = runs.get(1), exact = runs.get(4);
|
||||
assertThat(baseline.calls.get()).isEqualTo(600);
|
||||
assertThat(exact.hits).isEqualTo(exactRepeats);
|
||||
assertThat(at80.hits).isGreaterThan(exact.hits);
|
||||
assertThat(runs.get(0).wrong).isGreaterThanOrEqualTo(at80.wrong);
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------- 5. the traps
|
||||
|
||||
@Test
|
||||
void lookAlikesThatShareWordsButNotMeaning() {
|
||||
String[][] pairs = {
|
||||
{"How do I cancel my order?", "How do I keep my order and not cancel it?"},
|
||||
{"Is shipping free for orders over $50?", "Is shipping free for orders over $500?"},
|
||||
{"Do you ship to Canada?", "Do you ship to Cuba?"},
|
||||
{"What is the warranty on the laptop?", "What is the warranty on the monitor?"},
|
||||
{"Where is order 1001?", "Where is order 2002?"},
|
||||
{"Can I get a refund within 30 days?", "Can I get a refund after 30 days?"},
|
||||
{"I want to delete my account", "I do not want to delete my account"},
|
||||
};
|
||||
try (var t = new Transcript("05-lookalikes.txt", "Pairs that read alike and mean different things (cosine similarity)")) {
|
||||
t.line("%-45s | %-45s | %6s | %s", "cached question", "new question", "cosine", "Spring AI score = (1 + cosine) / 2");
|
||||
for (String[] p : pairs) {
|
||||
double c = Embedder.cosine(embeddings.embed(p[0]), embeddings.embed(p[1]));
|
||||
double s = (1 + c) / 2;
|
||||
t.line("%-45s | %-45s | %6.3f | %.3f %s", p[0], p[1], c, s, s >= 0.90 ? "<- served at threshold 0.90 and 0.80" : s >= 0.80 ? "<- served at threshold 0.80" : "");
|
||||
assertThat(s).isBetween(0.0, 1.0);
|
||||
}
|
||||
}
|
||||
long servedAt90 = java.util.Arrays.stream(pairs).filter(p -> (1 + Embedder.cosine(embeddings.embed(p[0]), embeddings.embed(p[1]))) / 2 >= 0.90).count();
|
||||
assertThat(servedAt90).isGreaterThan(0);
|
||||
}
|
||||
|
||||
@Test
|
||||
void tenantsDoNotShareAnswers() {
|
||||
SemanticCache cache = freshCache(0.80);
|
||||
ScriptedModel model = new ScriptedModel(intentByText, 0);
|
||||
SemanticCacheAdvisor advisor = new SemanticCacheAdvisor(cache, 0);
|
||||
ChatClient client = ChatClient.builder(model).defaultAdvisors(advisor).build();
|
||||
String q = "How do I reset my account password?";
|
||||
client.prompt().user(q).advisors(a -> a.param(SemanticCacheAdvisor.TENANT, "shop-a")).call().content();
|
||||
var other = client.prompt().user(q).advisors(a -> a.param(SemanticCacheAdvisor.TENANT, "shop-b")).call().chatClientResponse();
|
||||
var same = client.prompt().user(q).advisors(a -> a.param(SemanticCacheAdvisor.TENANT, "shop-a")).call().chatClientResponse();
|
||||
// the same lookup without the tenant filter, to show what the filter prevents
|
||||
var unfiltered = cache.nearest("shop-a", q);
|
||||
try (var t = new Transcript("06-tenant-isolation.txt", "Same question, two tenants")) {
|
||||
t.line("shop-a asks \"%s\" -> model call (calls so far: 1)", q);
|
||||
t.line("shop-b asks the identical question -> %s (model calls so far: %d)",
|
||||
Boolean.TRUE.equals(other.context().get(SemanticCacheAdvisor.HIT)) ? "CACHE HIT (leak)" : "model call, no hit", model.calls.get());
|
||||
t.line("shop-a asks again -> %s", Boolean.TRUE.equals(same.context().get(SemanticCacheAdvisor.HIT)) ? "cache hit" : "model call");
|
||||
t.line("entries in the index: %d (one per tenant)", cache.size());
|
||||
t.line("The filter is a tag on the document and an expression on the search: tenant == 'shop-b'.");
|
||||
t.line("nearest(shop-a, q) without asking as shop-b still finds: %s", unfiltered.map(h -> "\"" + h.storedQuestion() + "\"").orElse("nothing"));
|
||||
}
|
||||
assertThat(Boolean.TRUE.equals(other.context().get(SemanticCacheAdvisor.HIT))).isFalse();
|
||||
assertThat(Boolean.TRUE.equals(same.context().get(SemanticCacheAdvisor.HIT))).isTrue();
|
||||
assertThat(cache.size()).isEqualTo(2);
|
||||
}
|
||||
|
||||
@Test
|
||||
void anExpiredEntryStopsBeingServed() throws Exception {
|
||||
SemanticCache cache = freshCache(0.80);
|
||||
String id = cache.put(TENANT, "How do I reset my account password?", ScriptedModel.answerFor("reset-password"));
|
||||
boolean before = cache.lookup(TENANT, "How do I reset my account password?").isPresent();
|
||||
long sizeBefore = cache.size();
|
||||
cache.expireAfterSeconds(id, 1);
|
||||
Thread.sleep(1500);
|
||||
boolean after = cache.lookup(TENANT, "How do I reset my account password?").isPresent();
|
||||
try (var t = new Transcript("07-expiry.txt", "Expiry with EXPIRE on the document's key")) {
|
||||
t.line("stored one entry, index holds %d document(s); lookup -> %s", sizeBefore, before ? "hit" : "miss");
|
||||
t.line("EXPIRE sc:<id> 1, wait 1.5 s; index now holds %d document(s); lookup -> %s", cache.size(), after ? "hit" : "miss");
|
||||
}
|
||||
assertThat(before).isTrue();
|
||||
assertThat(after).isFalse();
|
||||
assertThat(cache.size()).isZero();
|
||||
}
|
||||
|
||||
@Test
|
||||
void anotherEmbeddingModelIsAnotherIndex() {
|
||||
SemanticCache cache = freshCache(0.80);
|
||||
cache.put(TENANT, "How do I reset my account password?", "a");
|
||||
EmbeddingModel eight = new EmbeddingModel() {
|
||||
@Override
|
||||
public org.springframework.ai.embedding.EmbeddingResponse call(org.springframework.ai.embedding.EmbeddingRequest r) {
|
||||
List<org.springframework.ai.embedding.Embedding> out = new ArrayList<>();
|
||||
for (int i = 0; i < r.getInstructions().size(); i++) out.add(new org.springframework.ai.embedding.Embedding(embed(r.getInstructions().get(i)), i));
|
||||
return new org.springframework.ai.embedding.EmbeddingResponse(out);
|
||||
}
|
||||
|
||||
@Override
|
||||
public float[] embed(org.springframework.ai.document.Document d) {
|
||||
return new float[8];
|
||||
}
|
||||
|
||||
@Override
|
||||
public float[] embed(String text) {
|
||||
float[] v = new float[8];
|
||||
v[Math.abs(text.hashCode()) % 8] = 1f;
|
||||
return v;
|
||||
}
|
||||
|
||||
@Override
|
||||
public List<float[]> embed(List<String> texts) {
|
||||
return texts.stream().map(this::embed).toList();
|
||||
}
|
||||
|
||||
@Override
|
||||
public int dimensions() { return 8; }
|
||||
};
|
||||
String outcome;
|
||||
try {
|
||||
// same index name and prefix, a model with 8 dimensions instead of 384; afterPropertiesSet() finds the index already exists
|
||||
var swapped = new org.springframework.ai.vectorstore.redis.RedisVectorStore[1];
|
||||
swapped[0] = org.springframework.ai.vectorstore.redis.RedisVectorStore.builder(jedis, eight)
|
||||
.indexName(SemanticCache.INDEX).prefix(SemanticCache.PREFIX)
|
||||
.metadataFields(org.springframework.ai.vectorstore.redis.RedisVectorStore.MetadataField.tag("tenant"), org.springframework.ai.vectorstore.redis.RedisVectorStore.MetadataField.text("answer"))
|
||||
.initializeSchema(true).build();
|
||||
swapped[0].afterPropertiesSet();
|
||||
swapped[0].add(List.of(new org.springframework.ai.document.Document("x", "hello", Map.of("tenant", TENANT, "answer", "b"))));
|
||||
var found = swapped[0].similaritySearch(org.springframework.ai.vectorstore.SearchRequest.builder().query("hello").topK(1)
|
||||
.filterExpression("tenant == '" + TENANT + "'").build());
|
||||
outcome = "no exception; search returned " + found.size() + " document(s): " + found.stream().map(d -> d.getText()).toList();
|
||||
} catch (Exception e) {
|
||||
Throwable root = e;
|
||||
while (root.getCause() != null) root = root.getCause();
|
||||
outcome = "exception " + root.getClass().getSimpleName() + ": " + root.getMessage();
|
||||
}
|
||||
try (var t = new Transcript("08-embedding-model-change.txt", "An index built for 384 dimensions, written by a model with 8")) {
|
||||
t.line("index created by MiniLM (384 dimensions); a new application version configures an 8-dimension model against the same index name.");
|
||||
t.line("add + search -> %s", outcome);
|
||||
}
|
||||
assertThat(outcome).isNotBlank();
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------- 9. where the milliseconds go
|
||||
|
||||
static double pct(List<Double> v, double p) {
|
||||
List<Double> s = v.stream().sorted().toList();
|
||||
return s.get(Math.min(s.size() - 1, (int) Math.ceil(p * s.size()) - 1));
|
||||
}
|
||||
|
||||
@Test
|
||||
void whereTheTimeGoesOnAHit() {
|
||||
SemanticCache cache = seededCache(0.80);
|
||||
List<String> questions = new ArrayList<>();
|
||||
data.paraphrases().forEach(r -> questions.add(r.text()));
|
||||
data.offTopic().forEach(r -> questions.add(r.text()));
|
||||
for (int i = 0; i < 10; i++) embeddings.embed(questions.get(i)); // warm-up, not measured
|
||||
List<Double> embed = new ArrayList<>(), search = new ArrayList<>();
|
||||
for (String q : questions) {
|
||||
long t0 = System.nanoTime();
|
||||
embeddings.embed(q);
|
||||
long t1 = System.nanoTime();
|
||||
cache.lookup(TENANT, q); // embeds again inside the store, then searches
|
||||
long t2 = System.nanoTime();
|
||||
embed.add((t1 - t0) / 1e6);
|
||||
search.add((t2 - t1) / 1e6 - (t1 - t0) / 1e6);
|
||||
}
|
||||
try (var t = new Transcript("09-hit-latency.txt", "Where the milliseconds go on a lookup (110 questions, 30 entries cached, 2 vCPUs)")) {
|
||||
t.line("embed the question (MiniLM in-process) : median %.1f ms, p95 %.1f ms", pct(embed, 0.5), pct(embed, 0.95));
|
||||
t.line("Redis KNN search with the tenant filter : median %.1f ms, p95 %.1f ms (lookup time minus one embedding)", pct(search, 0.5), pct(search, 0.95));
|
||||
t.line("Timings drift run to run; the shape (embedding dominates, Redis is small) does not.");
|
||||
}
|
||||
assertThat(pct(embed, 0.5)).isGreaterThan(pct(search, 0.5));
|
||||
}
|
||||
|
||||
@Test
|
||||
void againstARealLocalModel() throws Exception {
|
||||
org.junit.jupiter.api.Assumptions.assumeTrue(isUp(11555), "no Ollama on 127.0.0.1:11555");
|
||||
var api = org.springframework.ai.ollama.api.OllamaApi.builder().baseUrl("http://127.0.0.1:11555").build();
|
||||
var model = org.springframework.ai.ollama.OllamaChatModel.builder().ollamaApi(api)
|
||||
.options(org.springframework.ai.ollama.api.OllamaChatOptions.builder().model("tl").temperature(0.0).numPredict(60).build()).build();
|
||||
SemanticCache cache = freshCache(0.80);
|
||||
SemanticCacheAdvisor advisor = new SemanticCacheAdvisor(cache, 0);
|
||||
ChatClient client = ChatClient.builder(model).defaultAdvisors(advisor).build();
|
||||
client.prompt().user("Say hello").call().content(); // load the model, not measured
|
||||
cache = freshCache(0.80);
|
||||
advisor = new SemanticCacheAdvisor(cache, 0);
|
||||
client = ChatClient.builder(model).defaultAdvisors(advisor).build();
|
||||
|
||||
List<Dataset.Row> seeds = data.seeds().subList(0, 6);
|
||||
List<Dataset.Row> paras = new ArrayList<>();
|
||||
for (var s : seeds) paras.add(data.paraphrases().stream().filter(p -> p.intent().equals(s.intent())).findFirst().orElseThrow());
|
||||
List<Double> missMs = new ArrayList<>(), hitMs = new ArrayList<>();
|
||||
int wrongHits = 0;
|
||||
try (var t = new Transcript("10-real-model-latency.txt", "A real local model (TinyLlama 1.1B, Q4_0, 60 tokens) behind the advisor, 2 vCPUs")) {
|
||||
for (var s : seeds) {
|
||||
long t0 = System.nanoTime();
|
||||
var resp = client.prompt().user(s.text()).call().chatClientResponse();
|
||||
double ms = (System.nanoTime() - t0) / 1e6;
|
||||
boolean hit = Boolean.TRUE.equals(resp.context().get(SemanticCacheAdvisor.HIT));
|
||||
if (hit) {
|
||||
t.line("first ask %6.0f ms CACHE \"%s\" <- served the answer to \"%s\" (score %.3f), a DIFFERENT question", ms, s.text(),
|
||||
resp.context().get(SemanticCacheAdvisor.MATCHED), (Double) resp.context().get(SemanticCacheAdvisor.SCORE));
|
||||
wrongHits++;
|
||||
} else {
|
||||
missMs.add(ms);
|
||||
t.line("first ask %6.0f ms model \"%s\"", ms, s.text());
|
||||
}
|
||||
}
|
||||
for (var p : paras) {
|
||||
long t0 = System.nanoTime();
|
||||
var resp = client.prompt().user(p.text()).call().chatClientResponse();
|
||||
double ms = (System.nanoTime() - t0) / 1e6;
|
||||
boolean hit = Boolean.TRUE.equals(resp.context().get(SemanticCacheAdvisor.HIT));
|
||||
if (hit) hitMs.add(ms);
|
||||
t.line("reworded %6.0f ms %s \"%s\"%s", ms, hit ? "CACHE " : "model ", p.text(),
|
||||
hit ? " <- " + resp.context().get(SemanticCacheAdvisor.MATCHED) : "");
|
||||
}
|
||||
t.blank();
|
||||
t.line("median model call %.0f ms; median cache hit %.0f ms (%d of %d rewordings hit); first-time questions wrongly answered from the cache: %d of %d",
|
||||
pct(missMs, 0.5), hitMs.isEmpty() ? 0 : pct(hitMs, 0.5), hitMs.size(), paras.size(), wrongHits, seeds.size());
|
||||
}
|
||||
assertThat(hitMs).isNotEmpty();
|
||||
assertThat(pct(hitMs, 0.5)).isLessThan(pct(missMs, 0.5));
|
||||
}
|
||||
|
||||
static boolean isUp(int port) {
|
||||
try (var s = new java.net.Socket("127.0.0.1", port)) {
|
||||
return true;
|
||||
} catch (Exception e) {
|
||||
return false;
|
||||
}
|
||||
}
|
||||
}
|
||||
Reference in New Issue
Block a user