Exa policy guardrails
'Implement content policy enforcement, domain filtering, and usage guardrails for Exa.From its SKILL.md
npx -y skills add jeremylongshore/claude-code-plugins-plus-skills --skill exa-policy-guardrailsAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
What its file declares
Copied from the file, not written here
The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.
SKILL.md
6.3 KB, ~1.5k tokens by cl100k_base, as published. Nobody here has run it
Exa Policy Guardrails
Overview
Policy enforcement for Exa neural search integrations. Exa searches the open web, so results may include unreliable sources, competitor content, or inappropriate material. This skill covers domain allowlists/blocklists (via Exa's includeDomains/excludeDomains), content moderation, query sanitization, freshness policies, and per-user budget enforcement.
Prerequisites
exa-jsinstalled and configured- Content policy requirements defined
- Redis for per-user quota tracking (optional)
Instructions
Step 1: Domain Filtering (Built-in Exa Feature)
import Exa from "exa-js";
const exa = new Exa(process.env.EXA_API_KEY);
// Exa supports up to 1200 domains in includeDomains/excludeDomains
const TRUSTED_SOURCES = {
medical: [
"pubmed.ncbi.nlm.nih.gov", "who.int", "cdc.gov",
"nejm.org", "nature.com", "thelancet.com",
],
technical: [
"github.com", "stackoverflow.com", "developer.mozilla.org",
"docs.python.org", "nodejs.org", "arxiv.org",
],
news: [
"reuters.com", "apnews.com", "bbc.com",
"techcrunch.com", "arstechnica.com",
],
};
const BLOCKED_DOMAINS = [
"competitor1.com", "competitor2.io",
"spam-farm.com", "content-mill.net",
];
async function policySearch(
query: string,
category: keyof typeof TRUSTED_SOURCES | "general"
) {
const opts: any = {
type: "auto",
numResults: 10,
text: { maxCharacters: 1000 },
moderation: true, // Exa's built-in content moderation
};
if (category !== "general" && TRUSTED_SOURCES[category]) {
opts.includeDomains = TRUSTED_SOURCES[category];
} else {
opts.excludeDomains = BLOCKED_DOMAINS;
}
return exa.searchAndContents(query, opts);
}
Step 2: Query Content Policy
const BLOCKED_PATTERNS = [
/how to (hack|exploit|attack|ddos)/i,
/(buy|purchase|order)\s+(drugs|weapons|firearms)/i,
/personal.*(address|phone|ssn|social security)/i,
/generate.*(malware|ransomware|virus)/i,
];
function validateQuery(input: string): string {
for (const pattern of BLOCKED_PATTERNS) {
if (pattern.test(input)) {
throw new PolicyViolation("Query blocked by content policy");
}
}
// Sanitize
return input
.replace(/[<>{}]/g, "") // strip HTML/template chars
.replace(/\0/g, "") // remove null bytes
.trim()
.substring(0, 500); // cap query length
}
class PolicyViolation extends Error {
constructor(message: string) {
super(message);
this.name = "PolicyViolation";
}
}
Step 3: Freshness Policy
// Enforce minimum recency for time-sensitive use cases
function applyFreshnessPolicy(
opts: any,
maxAgeDays: number
): any {
const cutoff = new Date(Date.now() - maxAgeDays * 24 * 60 * 60 * 1000);
return {
...opts,
startPublishedDate: cutoff.toISOString(),
};
}
// Usage: only return results from the last 90 days
const results = await exa.searchAndContents("AI regulation updates",
applyFreshnessPolicy(
{ type: "neural", numResults: 10, text: true },
90 // max 90 days old
)
);
Step 4: Per-User Budget Enforcement
class ExaUsagePolicy {
private usage = new Map<string, { count: number; resetAt: number }>();
private limits: Record<string, number>;
constructor(limits: Record<string, number> = {
"free": 10,
"pro": 100,
"enterprise": 1000,
}) {
this.limits = limits;
}
checkQuota(userId: string, tier: string): void {
const limit = this.limits[tier] || this.limits["free"] || 10;
const now = Date.now();
const hourKey = `${userId}:${new Date().toISOString().substring(0, 13)}`;
let entry = this.usage.get(hourKey);
if (!entry || entry.resetAt < now) {
entry = { count: 0, resetAt: now + 3600 * 1000 };
}
if (entry.count >= limit) {
throw new PolicyViolation(
`Hourly search quota exceeded: ${entry.count}/${limit}`
);
}
entry.count++;
this.usage.set(hourKey, entry);
}
}
const usagePolicy = new ExaUsagePolicy();
Step 5: Combined Policy Enforcement
async function enforcedSearch(
userId: string,
userTier: string,
rawQuery: string,
category: keyof typeof TRUSTED_SOURCES | "general" = "general",
maxAgeDays?: number
) {
// 1. Check quota
usagePolicy.checkQuota(userId, userTier);
// 2. Validate and sanitize query
const query = validateQuery(rawQuery);
// 3. Build options with domain policy
let opts: any = {
type: "auto",
numResults: 10,
text: { maxCharacters: 1000 },
moderation: true,
};
if (category !== "general" && TRUSTED_SOURCES[category]) {
opts.includeDomains = TRUSTED_SOURCES[category];
} else {
opts.excludeDomains = BLOCKED_DOMAINS;
}
// 4. Apply freshness policy
if (maxAgeDays) {
opts = applyFreshnessPolicy(opts, maxAgeDays);
}
// 5. Execute search
return exa.searchAndContents(query, opts);
}
Error Handling
| Issue | Cause | Solution |
|---|---|---|
| Competitor content in results | No domain filtering | Apply excludeDomains blocklist |
| Harmful query accepted | No content policy | Validate queries against blocked patterns |
| Stale results displayed | No freshness check | Apply startPublishedDate filter |
| API cost overrun | No usage limits | Implement per-user/tier quotas |
| Blocked policy query | False positive | Review and adjust BLOCKED_PATTERNS |
Resources
Next Steps
For architecture decisions, see exa-architecture-variants. For cost control, see exa-cost-tuning.
What ships with it
Read from the repository
Just SKILL.md. No reference files, no scripts.