Building a relevance ai Workflow for Lead Enrichment

If you’ve ever tried to enrich a lead list with context from the web, you know how quickly manual research eats up your day. After testing a few approaches, I found that a relevance ai pipeline can pull in the right snippets, score them, and dump the results into a spreadsheet—all without you lifting a finger.

Why relevance ai matters for solo operators

solo operators wear many hats. You spend time prospecting, crafting outreach, and chasing replies. When you add manual research to that mix, the day disappears. A relevance ai pipeline changes that by automating the lookup step. It takes a lead’s name or domain, queries a search API, and returns the most relevant snippet with a confidence score. You then decide whether to keep, enrich, or discard the lead.

The beauty is that you don’t need a team of data scientists. You just need a prompt that tells the model what “relevant” means for your niche. Once the prompt is stable, the pipeline runs on a schedule and feeds your CRM.

How do you keep relevance ai prompts from drifting off-topic?

One of the first things I noticed is that a vague prompt makes the model wander. It starts returning news articles about unrelated companies or generic definitions. To keep it focused, I anchor the prompt with three constraints: the industry, the type of information I want, and the format of the output.

Here’s the prompt I use for a SaaS lead enrichment task:

• Given the company name {{name}} and domain {{domain}}, return a one‑sentence description of its core product, the primary problem it solves for mid‑market teams, and a confidence score between 0 and 1. If you cannot find reliable information, return “unknown” and a score of 0.

The double curly braces are placeholders that my automation fills in before calling the relevance ai endpoint. By fixing the output shape, I reduce hallucinations and make downstream parsing trivial.

What most guides get wrong about relevance ai tuning

Many tutorials tell you to increase the temperature setting to get more creative answers. That advice backfires when you need factual consistency. A higher temperature makes the model invent details that look plausible but are wrong. I keep the temperature at 0.2 and rely on the prompt to enforce specificity.

Another common mistake is treating the confidence score as a guarantee. The score is a probability, not a truth value. I use it as a filter: anything below 0.6 gets flagged for manual review. This simple threshold cut my false‑positive rate by almost half.

Finally, guides often ignore rate limits. They show a single call and assume you can scale linearly. In reality, the vendor enforces a hard limit of 150 requests per minute on the basic tier. If you blast through a list of 500 leads, you’ll get 429 errors and stall the pipeline.

Concrete named example: building a lead enrichment pipeline with relevance ai

Let’s walk through a real build I ran last month for a freelance outreach agency. The stack consisted of three pieces: a Google Sheet for input, a Make (formerly Integromat) scenario to orchestrate calls, and the relevance ai API for the enrichment.

First, I created a sheet with columns: Name, Domain, Enrichment, Score, Status. The sheet holds up to 2000 rows, which is enough for a month’s worth of prospects.

Next, I built a Make.com scenario that watches for new rows. When a row appears, it pulls the Name and Domain, inserts them into the prompt template, and POSTs to the relevance ai endpoint. The endpoint returns JSON with fields text and score. I then write those values back to the Enrichment and Score columns.

Here’s the exact HTTP request I used (you can copy it into Make’s HTTP module):

• POST https://api.relevanceai.com/v1/enrich
Headers: Authorization: Bearer {{your_api_key}}, Content-Type: application/json
Body: {
"prompt": "Given the company name {{Name}} and domain {{Domain}}, return a one‑sentence description of its core product, the primary problem it solves for mid‑market teams, and a confidence score between 0 and 1. If you cannot find reliable information, return \"unknown\" and a score of 0.",
"temperature": 0.2,
"max_tokens": 150
}

The scenario runs every five minutes. If the score is below 0.6, the Status column gets set to “review”; otherwise it’s marked “done”. I then filter the sheet for “done” rows and export them to my outreach sequencer.

Cost wise, the relevance ai API charges $0.008 per call on the basic tier. At 150 calls per minute, that’s $7.20 per hour if you run continuously. For my typical usage of 3000 calls a day, the monthly bill lands around $18. I added a $5 buffer for overhead, so I pay roughly $23/mo.

How to debug when relevance ai returns junk

Even with a tight prompt, you’ll occasionally see garbage output. The first thing I check is the actual prompt that was sent. Logging the raw request in Make.com helps me see if the placeholders were filled correctly. A missing domain or a typo in the variable name will cause the model to hallucinate.

If the prompt looks fine, I look at the temperature. A value above 0.5 often produces creative but inaccurate answers. Dropping it back to 0.2 usually clears things up.

Next, I inspect the confidence score. A score of 0.1 or 0.2 indicates the model didn’t find strong matches. In that case, I broaden the search by removing the “one‑sentence” constraint and asking for a short paragraph instead. That sometimes pulls in a relevant snippet that the stricter format missed.

Finally, I check the API’s error codes. A 429 means you’ve hit the rate limit. The fix is to add a delay between batches or upgrade to a higher tier. A 500 means the service had a hiccup; retrying after a few seconds normally works.

One concrete gripe I have: the vendor’s documentation hides the rate limit behind a FAQ page, and I burned through my quota on a test run before I even realized there was a cap. It’s annoying, but once you know the limit you can plan around it.

On the love side, I really appreciate that the relevance ai endpoint returns a confidence score with every response. That single number lets me automate quality control without building a separate model.

Pricing, love, and gripes

The basic tier costs $29 per month and includes 150 requests per minute. For a solo operator doing a few thousand enrichment calls a day, that’s fair. The pro tier jumps to $99 per month and raises the limit to 500 requests per minute. Unless you’re pushing tens of thousands of calls, the pro tier feels steep for what you get.

I think the $49/mo “mid” plan is overpriced—it only adds a marginally higher concurrency limit that most solo users never hit. That opinion could be wrong if you run a team of five or more.

Aside from the rate‑limit surprise, the only other hiccup I’ve seen is occasional latency spikes during peak hours. The response time can jump from 300 ms to over a second, which makes the Make.com scenario timeout if you don’t increase the retry interval. — and good luck finding docs for this — you’ll need to adjust the timeout setting in the module.

Adjacent reading: deeper coverage of AI agent platforms.

If you’d rather skip the build and deploy a working version in an afternoon, we’ve packaged this workflow as a blueprint at deepusecase.com/vault.