rene/transceiver-db
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: 100% verified data — no invented prices, part numbers, or designations

Browse Source 
gatherBlogData():
- Fetches real prices from price_observations (last 30 days) per product
- Filters transceivers by speed extracted from topic keywords
- Enriches every product with verified_prices array + has_verified_price flag
- Joins DB products with vector search results (DB first — they have real prices)

contextData injection (blog.ts):
- [PRODUCT] lines: exact standard_name, form_factor, speed, reach, connector, dBm specs, Watts
- [VERIFIED PRICE] lines: real EUR/USD price, vendor, observed date, source URL
- [NO VERIFIED PRICE IN DB]: explicit tag — LLM must not invent a number
- [NO PRODUCT DATA AVAILABLE]: fallback when DB returns nothing

fo-blog-pipeline.ts system prompt:
- DATA INTEGRITY RULES block: prices/part numbers/vendors ONLY from context
- Never approximate with ~€350 or 'typically $200-600' for specific products
- Power specs only from [PRODUCT] data or REFERENCE VALUES

STEP4 context instructions:
- Explicit rules on how to use [VERIFIED PRICE] vs [NO VERIFIED PRICE]
- Invented data = HARD FAIL in QA

STEP9 QA — 3 new hard fail checks (30, 31, 32):
- Check 30: invented prices → remove or replace with flexoptix.net reference
- Check 31: invented part numbers → remove, use class name instead
- Check 32: invented vendor names → remove if not in known list
This commit is contained in:
Rene Fichtmueller 2026-04-01 17:27:55 +02:00
parent ee8b3c0779
commit 174078efdb
2 changed files with 170 additions and 37 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

31
packages/api/src/llm/fo-blog-pipeline.ts
View File

@ -91,6 +91,15 @@ BANNED STRUCTURAL PATTERNS:
- OEM pricing for compatible optics — "400G DR4 at $2,000-5,000" is OEM pricing. Compatible vendor range (Flexoptix, FS, ProLabs) is typically $200-600. Always specify OEM vs compatible. - OEM pricing for compatible optics — "400G DR4 at $2,000-5,000" is OEM pricing. Compatible vendor range (Flexoptix, FS, ProLabs) is typically $200-600. Always specify OEM vs compatible.
- Markdown headers (##, ###, ####, **bold headers**) anywhere in the article body — write in plain text. No hash symbols, no asterisk headers. Section titles as plain sentence or not at all. - Markdown headers (##, ###, ####, **bold headers**) anywhere in the article body — write in plain text. No hash symbols, no asterisk headers. Section titles as plain sentence or not at all.
DATA INTEGRITY RULES (ABSOLUTE — harder than anything else on this list):
- EVERY price, part number, and product designation in the article MUST come from the CONTEXT DATA block below, tagged [VERIFIED PRICE] or [PRODUCT].
- If a product has [NO VERIFIED PRICE IN DB], you MUST NOT write any price for it. Write "current pricing at flexoptix.net" instead.
- NEVER invent, estimate, or approximate a price. Not "~€350", not "around $400", not "typically $200-600 for compatible". Only real [VERIFIED PRICE] values from the context.
- NEVER invent a part number. If you don't see it in [PRODUCT] lines, don't use it.
- NEVER invent a vendor. Only use vendor names from the [PRODUCT] or [VERIFIED PRICE] lines.
- If the context has no products at all ([NO PRODUCT DATA AVAILABLE]), write the article without any specific product names or prices — use technology class names only (e.g., "400G DR4 optics" not "the Flexoptix FX-400DR4-001").
- Power specs (dBm, Watts) may ONLY be cited if they appear in the [PRODUCT] data or in the REFERENCE VALUES section below. Never derive or estimate them.
HARD RULES (non-negotiable — article FAILS QA without these): HARD RULES (non-negotiable — article FAILS QA without these):
1. Start with a BRUTAL hook — not "If you're still..." but "You're about to sign a PO. Stop." 1. Start with a BRUTAL hook — not "If you're still..." but "You're about to sign a PO. Stop."
2. Include a "WHAT BREAKS IN PRODUCTION" section with at least 2 SPECIFIC failure scenarios: 2. Include a "WHAT BREAKS IN PRODUCTION" section with at least 2 SPECIFIC failure scenarios:
@ -270,10 +279,18 @@ FORMATTING RULES (both styles — non-negotiable):
MINIMUM 2000 words. No placeholders. No TODO markers. Complete article. MINIMUM 2000 words. No placeholders. No TODO markers. Complete article.
CONTEXT DATA RULES (read before writing a single word):
The context below contains REAL data from the Flexoptix product database.
- Lines starting with [PRODUCT] = real products. Use their exact names, form factors, speeds.
- Lines starting with [VERIFIED PRICE] = real prices observed in our DB. You MAY cite these.
- Lines with [NO VERIFIED PRICE IN DB] = no price data exists. Write "see flexoptix.net for current pricing" — NEVER invent a number.
- Lines with [NO PRODUCT DATA AVAILABLE] = DB returned nothing. Use technology class names only.
- ANY price, part number, or product name you write that is NOT in the context below is an invention. Inventions are HARD FAILS in QA.
Outline: Outline:
{{OUTLINE}} {{OUTLINE}}
Context data: Context data from Flexoptix database (verified — use exactly as provided):
{{CONTEXT_DATA}}`; {{CONTEXT_DATA}}`;
// ═══════════════════════════════════════════════════════ // ═══════════════════════════════════════════════════════
@ -568,6 +585,18 @@ CALIBRATION FAILS (auto-reject — fix before returning):
→ STRONG: "You're about to spend $400,000 on optics. Here's how to accidentally turn it into a $2M problem." → STRONG: "You're about to spend $400,000 on optics. Here's how to accidentally turn it into a $2M problem."
→ If the hook lacks a concrete number or consequence, strengthen it. → If the hook lacks a concrete number or consequence, strengthen it.
30. INVENTED PRICES (HARD FAIL): Check EVERY price mentioned in the article.
→ Any price that was NOT in the [VERIFIED PRICE] lines of the context data is invented.
→ REMOVE invented prices. Replace with "see flexoptix.net for current pricing" or a technology-class range from the REFERENCE VALUES section.
→ Exception: general ranges like "$200-600 for compatible 400G DR4" from the system prompt reference values are acceptable ONLY if no verified price exists for a specific product.
→ If an exact price like "€312.50" or "$449.99" appears and it was NOT in the context — REMOVE IT.
31. INVENTED PART NUMBERS (HARD FAIL): Check every part number, SKU, or model number.
→ If it was NOT in the [PRODUCT] lines of the context data, it is invented.
→ REMOVE invented part numbers. Replace with the product class name (e.g., "400G DR4 optic" not "FX-QSFPDD-400G-DR4-001").
32. INVENTED VENDOR NAMES: Any vendor cited that was NOT in the context data or in the system prompt reference list (Cisco, Juniper, Arista, Flexoptix, FS.com, ProLabs, InnoLight, Coherent, Lumentum) — REMOVE.
CRITICAL OUTPUT RULE: CRITICAL OUTPUT RULE:
Return ONLY the fixed article text. NO review commentary. NO numbered issue lists. NO "Critical Review" section. NO "HARD FAIL CHECKS" header. NO markdown review structure. Return ONLY the fixed article text. NO review commentary. NO numbered issue lists. NO "Critical Review" section. NO "HARD FAIL CHECKS" header. NO markdown review structure.

176
packages/api/src/routes/blog.ts
View File

@ -183,7 +183,10 @@ const BLOG_TEMPLATES: Record<string, BlogTemplate[]> = {
}; };
/** Gather data from vector collections for blog content — with PostgreSQL fallback. /** Gather data from vector collections for blog content — with PostgreSQL fallback.
* Topic-aware: strategy articles (hype_cycle, comparison) skip troubleshooting data. */ * Topic-aware: strategy articles (hype_cycle, comparison) skip troubleshooting data.
*
* IMPORTANT: Always enriches products with REAL verified prices from price_observations.
* The LLM may ONLY use prices returned here — never invent pricing. */
async function gatherBlogData(keywords: string[], topic?: string): Promise<{ async function gatherBlogData(keywords: string[], topic?: string): Promise<{
products: Array<Record<string, unknown>>; products: Array<Record<string, unknown>>;
news: Array<Record<string, unknown>>; news: Array<Record<string, unknown>>;
@ -195,40 +198,33 @@ async function gatherBlogData(keywords: string[], topic?: string): Promise<{
// Strategy articles should NOT pull troubleshooting data (topic separation) // Strategy articles should NOT pull troubleshooting data (topic separation)
const skipTroubleshooting = topic === "hype_cycle" || topic === "comparison" || topic === "new_product"; const skipTroubleshooting = topic === "hype_cycle" || topic === "comparison" || topic === "new_product";
// Try vector search first (requires Qdrant + embeddings) // Extract speed/form_factor hints from keywords for relevance filtering
try { const speedHints = keywords.join(" ").match(/\b(10|25|40|100|200|400|800|1600)G\b/gi) || [];
const [products, news, faq, troubleshooting] = await Promise.all([ const speedGbps = speedHints.map(s => parseInt(s)).filter(Boolean);
semanticSearch("product_embeddings", query, 10).catch(() => []),
semanticSearch("news_embeddings", query, 5).catch(() => []),
skipTroubleshooting ? Promise.resolve([]) : semanticSearch("faq_embeddings", query, 5).catch(() => []),
skipTroubleshooting ? Promise.resolve([]) : semanticSearch("troubleshooting_embeddings", query, 3).catch(() => []),
]);
const result = { // ── Fetch real products with verified prices from DB ──────────────────────
products: products.map((r) => ({ score: r.score, ...r.payload })), // Primary: filter by keyword-extracted speed; fallback to top products by speed
news: news.map((r) => ({ score: r.score, ...r.payload })), const productQuery = speedGbps.length > 0
faq: faq.map((r) => ({ score: r.score, ...r.payload })), ? `SELECT t.id, t.slug, t.form_factor, t.speed, t.speed_gbps, t.reach_label,
troubleshooting: troubleshooting.map((r) => ({ score: r.score, ...r.payload })), t.fiber_type, t.standard_name, t.connector, t.power_consumption_w,
}; t.tx_power_min_dbm, t.tx_power_max_dbm, t.rx_sensitivity_dbm,
v.name as vendor, v.type as vendor_type
// If we got data from vector search, return it FROM transceivers t
if (result.products.length > 0 || result.news.length > 0) { LEFT JOIN vendors v ON t.vendor_id = v.id
return result; WHERE t.speed_gbps = ANY($1::int[])
} ORDER BY v.type = 'Compatible' DESC, t.speed_gbps DESC
} catch { LIMIT 20`
console.log("Vector search unavailable, falling back to PostgreSQL"); : `SELECT t.id, t.slug, t.form_factor, t.speed, t.speed_gbps, t.reach_label,
} t.fiber_type, t.standard_name, t.connector, t.power_consumption_w,
t.tx_power_min_dbm, t.tx_power_max_dbm, t.rx_sensitivity_dbm,
// Fallback: query PostgreSQL directly for product and news data v.name as vendor, v.type as vendor_type
const [productsDb, newsDb] = await Promise.all([
pool.query(
`SELECT t.slug, t.form_factor, t.speed, t.reach_label, t.fiber_type, t.standard_name,
v.name as vendor
FROM transceivers t FROM transceivers t
LEFT JOIN vendors v ON t.vendor_id = v.id LEFT JOIN vendors v ON t.vendor_id = v.id
ORDER BY t.speed_gbps DESC ORDER BY t.speed_gbps DESC
LIMIT 15` LIMIT 20`;
).catch(() => ({ rows: [] })),
const [productsDb, newsDb] = await Promise.all([
pool.query(productQuery, speedGbps.length > 0 ? [speedGbps] : []).catch(() => ({ rows: [] })),
pool.query( pool.query(
`SELECT title, source, category, published_at::text as date `SELECT title, source, category, published_at::text as date
FROM news_articles FROM news_articles
@ -237,8 +233,78 @@ async function gatherBlogData(keywords: string[], topic?: string): Promise<{
).catch(() => ({ rows: [] })), ).catch(() => ({ rows: [] })),
]); ]);
// ── Enrich each product with real verified prices ─────────────────────────
const productIds = productsDb.rows.map((r: Record<string, unknown>) => r.id).filter(Boolean);
let priceMap: Record<string, Array<{vendor: string; price: number; currency: string; url: string; observed_at: string}>> = {};
if (productIds.length > 0) {
const priceResult = await pool.query(
`SELECT po.transceiver_id,
v.name as vendor,
v.type as vendor_type,
po.price::float as price,
po.currency,
po.url,
po.time::text as observed_at
FROM price_observations po
JOIN vendors v ON po.source_vendor_id = v.id
WHERE po.transceiver_id = ANY($1::int[])
AND po.time > NOW() - INTERVAL '30 days'
AND po.price IS NOT NULL
AND po.price > 0
ORDER BY po.transceiver_id, po.time DESC`,
[productIds]
).catch(() => ({ rows: [] }));
// Group by transceiver_id — keep best price per vendor
for (const row of priceResult.rows) {
const tid = String(row.transceiver_id);
if (!priceMap[tid]) priceMap[tid] = [];
// Deduplicate by vendor — keep most recent
if (!priceMap[tid].find((p) => p.vendor === row.vendor)) {
priceMap[tid].push({
vendor: row.vendor,
price: row.price,
currency: row.currency || "EUR",
url: row.url || "",
observed_at: row.observed_at,
});
}
}
}
// Attach prices to products
const enrichedProducts = productsDb.rows.map((p: Record<string, unknown>) => ({
...p,
verified_prices: priceMap[String(p.id)] || [],
has_verified_price: (priceMap[String(p.id)] || []).length > 0,
}));
// Try vector search to supplement (but always use DB products as base — they have real prices)
try {
const [vectorProducts, news, faq, troubleshooting] = await Promise.all([
semanticSearch("product_embeddings", query, 10).catch(() => []),
semanticSearch("news_embeddings", query, 5).catch(() => []),
skipTroubleshooting ? Promise.resolve([]) : semanticSearch("faq_embeddings", query, 5).catch(() => []),
skipTroubleshooting ? Promise.resolve([]) : semanticSearch("troubleshooting_embeddings", query, 3).catch(() => []),
]);
return { return {
products: productsDb.rows, // DB products first (they have real prices) — vector results supplemental only
products: [
...enrichedProducts,
...vectorProducts.map((r) => ({ score: r.score, ...r.payload })),
].slice(0, 20),
news: news.length > 0 ? news.map((r) => ({ score: r.score, ...r.payload })) : newsDb.rows,
faq: faq.map((r) => ({ score: r.score, ...r.payload })),
troubleshooting: troubleshooting.map((r) => ({ score: r.score, ...r.payload })),
};
} catch {
// Vector search unavailable — use PostgreSQL only
}
return {
products: enrichedProducts,
news: newsDb.rows, news: newsDb.rows,
faq: [], faq: [],
troubleshooting: [], troubleshooting: [],
@ -968,10 +1034,48 @@ async function runLlmPipeline(
// Warmup // Warmup
await generate("Test", "OK", { temperature: 0.1, maxTokens: 8, timeoutMs: 60000 }).catch(() => {}); await generate("Test", "OK", { temperature: 0.1, maxTokens: 8, timeoutMs: 60000 }).catch(() => {});
// Build context data string for injection // Build context data string for injection — REAL DB data only, never fabricated
const contextData = data.products.slice(0, 15).map(p => type PriceEntry = { vendor: string; price: number; currency: string; url: string; observed_at: string };
`${p.standard_name || p.slug}: ${p.form_factor} ${p.speed}, reach ${p.reach_label || "N/A"}, fiber ${p.fiber_type || "N/A"}, vendor ${p.vendor || "N/A"}${p.price ? `, ~€${p.price}` : ""}` const contextLines: string[] = [];
).join("\n");
for (const p of data.products.slice(0, 20)) {
const prices = (p.verified_prices as PriceEntry[] | undefined) || [];
const hasPrice = prices.length > 0;
// Build product line with real specs
let line = `[PRODUCT] ${p.standard_name || p.slug || "unknown"}`;
if (p.form_factor) line += ` | Form factor: ${p.form_factor}`;
if (p.speed) line += ` | Speed: ${p.speed}`;
if (p.reach_label) line += ` | Reach: ${p.reach_label}`;
if (p.fiber_type) line += ` | Fiber: ${p.fiber_type}`;
if (p.connector) line += ` | Connector: ${p.connector}`;
if (p.vendor) line += ` | Vendor: ${p.vendor}`;
if (p.vendor_type) line += ` (${p.vendor_type})`;
// Optical specs if available
if (p.tx_power_min_dbm != null) line += ` | TX min: ${p.tx_power_min_dbm} dBm`;
if (p.tx_power_max_dbm != null) line += ` TX max: ${p.tx_power_max_dbm} dBm`;
if (p.rx_sensitivity_dbm != null) line += ` | RX sensitivity: ${p.rx_sensitivity_dbm} dBm`;
if (p.power_consumption_w != null) line += ` | Power: ${p.power_consumption_w}W`;
contextLines.push(line);
// Append verified prices — clearly tagged as real DB observations
if (hasPrice) {
for (const pr of prices.slice(0, 3)) {
const date = pr.observed_at ? pr.observed_at.split("T")[0] : "recent";
contextLines.push(
` [VERIFIED PRICE] ${pr.currency} ${pr.price.toFixed(2)} — ${pr.vendor} (observed ${date}) ${pr.url ? `| ${pr.url}` : ""}`
);
}
} else {
contextLines.push(` [NO VERIFIED PRICE IN DB — do NOT invent a price for this product]`);
}
}
const contextData = contextLines.length > 0
? contextLines.join("\n")
: "[NO PRODUCT DATA AVAILABLE — do NOT invent product names, part numbers, or prices]";
// Get blog type config // Get blog type config
const blogType = BLOG_TYPES[selectedTopic as keyof typeof BLOG_TYPES] || BLOG_TYPES.tutorial; const blogType = BLOG_TYPES[selectedTopic as keyof typeof BLOG_TYPES] || BLOG_TYPES.tutorial;