Tyler Gillispie
8f61cbdc49
- Renamed `forum/experiments` to `forum/poc-g1` to designate generation 1. - Created `forum/poc-g2` and a new `lab.ts` runner. - Non-destructively migrated `dag_engine_poc.ts`, `git_storage_poc.ts`, `merkle_diff_poc.ts`, and `frontmatter_poc.ts` to `poc-g2`. - Upgraded migrated PoCs to utilize actual production-ready tools (e.g. `std/yaml` parsing and isolated `Deno.Command` Git repos) per blueprint constraints. Co-authored-by: google-labs-jules[bot] <161369871+google-labs-jules[bot]@users.noreply.github.com> Co-authored-by: mrteye <1945243+mrteye@users.noreply.github.com>
85 lines
2.9 KiB
TypeScript
85 lines
2.9 KiB
TypeScript
/**
|
|
* Agent Forum v4 - Multi-Vec Isolation PoC
|
|
*
|
|
* Verifies the concept of preventing semantic bleed by utilizing isolated,
|
|
* separate vector databases for different domains (e.g., docs vs telemetry)
|
|
* rather than dumping all embeddings into a single database.
|
|
*/
|
|
|
|
// Mock representation of an embedded vector database instance
|
|
class MockVectorDB {
|
|
private data: Map<string, number[]> = new Map();
|
|
public name: string;
|
|
|
|
constructor(name: string) {
|
|
this.name = name;
|
|
}
|
|
|
|
insert(id: string, vector: number[]) {
|
|
this.data.set(id, vector);
|
|
}
|
|
|
|
// Simplified cosine similarity mock
|
|
query(vector: number[]): { id: string, score: number }[] {
|
|
const results = [];
|
|
for (const [id, vec] of this.data.entries()) {
|
|
// In a real scenario, this is mathematically calculating cosine similarity
|
|
// For the PoC, we just check if it's the exact same vector for a 1.0 score
|
|
const isExactMatch = vector.every((val, i) => val === vec[i]);
|
|
if (isExactMatch) {
|
|
results.push({ id, score: 1.0 });
|
|
} else {
|
|
// Mock random low score for non-matches
|
|
results.push({ id, score: 0.1 });
|
|
}
|
|
}
|
|
return results.sort((a, b) => b.score - a.score);
|
|
}
|
|
}
|
|
|
|
function runPoC() {
|
|
console.log("Running Multi-Vec Isolation PoC tests...");
|
|
|
|
// 1. Initialize isolated databases
|
|
const docsDb = new MockVectorDB("docs_graph.sqlite");
|
|
const telemetryDb = new MockVectorDB("telemetry_graph.sqlite");
|
|
|
|
// 2. Insert domain-specific data
|
|
// Mock vector for "How to implement authentication"
|
|
const authDocVector = [0.1, 0.8, 0.2];
|
|
docsDb.insert("doc_auth_guide", authDocVector);
|
|
|
|
// Mock vector for "High latency in database query"
|
|
const latencyTelemetryVector = [0.9, 0.1, 0.1];
|
|
telemetryDb.insert("tel_high_latency", latencyTelemetryVector);
|
|
|
|
// 3. Query the Docs DB for an architecture question
|
|
console.log(`Querying ${docsDb.name} for architecture context...`);
|
|
const docsResult = docsDb.query(authDocVector);
|
|
|
|
if (docsResult[0].id === "doc_auth_guide" && docsResult[0].score > 0.8) {
|
|
console.log(`✅ Found relevant doc in ${docsDb.name}`);
|
|
} else {
|
|
console.error(`❌ Failed to find doc in ${docsDb.name}`);
|
|
Deno.exit(1);
|
|
}
|
|
|
|
// 4. Prove Semantic Isolation (No Bleed)
|
|
// If we query the Telemetry DB with an architecture question, it should NOT return telemetry data
|
|
console.log(`Querying ${telemetryDb.name} with architecture context to prove isolation...`);
|
|
const isolatedResult = telemetryDb.query(authDocVector);
|
|
|
|
if (isolatedResult[0].score < 0.5) {
|
|
console.log(`✅ Semantic isolation confirmed. Telemetry DB did not return high confidence for a docs query.`);
|
|
} else {
|
|
console.error(`❌ Semantic bleed detected!`);
|
|
Deno.exit(1);
|
|
}
|
|
|
|
console.log("✅ Multi-Vec Isolation PoC successful: Domain-specific semantic bleed prevented.");
|
|
}
|
|
|
|
if (import.meta.main) {
|
|
runPoC();
|
|
}
|