- [Is scraping Google legal? What the 2026 court rulings say](https://parallel.ai/articles/is-scraping-google-legal)
Tags:Guides
Reading time: 10 minSeptember 26, 2026
# Exa Websets vs. Parallel FindAll API: A comprehensive comparison
Parallel FindAll and Exa Websets both turn a natural language query into a structured, enriched set of entities, and they take different architectural routes to get there. This comparison covers what each product does, how the two pipelines work, feature-by-feature differences, recall and accuracy, developer experience, and which use cases fit each.
Tags:Comparison
Reading time: 6 min## **What each product does**
### **Parallel FindAll API**
Parallel FindAll API is a web-scale entity discovery system that turns natural language queries into structured, enriched datasets. You describe what you're looking for in plain English ("Find all dental practices in Ohio with 4+ star Google reviews") and FindAll returns a structured table of matching entities, complete with citations and confidence scores.
FindAll is part of Parallel's broader Web Agent API suite, which also includes the Task API (for enrichment), Search API, Entity Search API (real-time company search with structured results in seconds), Responses API, and Monitor API. FindAll launched in late 2025 and is currently in public beta.
### **Exa Websets**
Exa Websets is Exa's high-compute web search product designed for complex queries that return hundreds or thousands of results. Built on top of Exa's embeddings-based search engine, Websets finds entities matching your criteria and then allows you to layer on enrichments: additional columns of data that an AI agent researches and fills in for each result.
Exa is used by companies like Databricks, Flatfile, OpenRouter, and StackAI for its core Search API. Cursor integrates Exa as a plugin in its marketplace. Websets is the newer, higher-level product built on that foundation.
## **Architecture and how they work**
### **Parallel FindAll: three-stage pipeline**
FindAll executes a structured three-stage pipeline:
**Candidate generation.** Searches Parallel's proprietary web index to identify potential entities matching your query. Unlike traditional search, it generates candidates dynamically based on your specific criteria.
**Match evaluation.** Each candidate is evaluated against your match conditions using multi-hop reasoning across web sources. Candidates that satisfy all conditions reach "matched" status; the rest are marked "unmatched."
**Structured enrichment.** Matched entities can be enriched with additional fields via Parallel's Task API. Enrichments can be added at creation time or retroactively, even on completed runs.
This separation between match conditions (boolean/filterable criteria that determine inclusion) and enrichments (arbitrary data extraction that doesn't affect matching) is a key architectural choice. You filter first, then enrich only the matches, which keeps costs down.
### **Exa Websets: search + enrichment columns**
Exa Websets works differently:
**Search.** A natural language query is processed through Exa's embeddings-based search engine, which finds web pages matching your criteria. Results are returned as items in a "Webset."
**Enrichment.** You define enrichment columns (text, date, number, options, email, phone, or URL) that an AI agent researches and populates for each item. Enrichments can be created via the API or through the dashboard's chat interface.
Exa's approach is more tightly coupled. There isn't an explicit "match evaluation" stage with multi-hop reasoning. The search engine finds candidates, and enrichments add data. Filtering happens primarily at the search level, with criteria used to verify matches.
## **Key feature comparison**
| Feature | Parallel FindAll | Exa Websets + Enrichments |
|---|---|---|
| Input | Natural language query | Natural language query |
| Entity discovery | Three-stage pipeline with explicit match evaluation | Embeddings-based search with criteria verification |
| Match conditions | Boolean criteria with multi-hop reasoning | Search query filters + criteria |
| Output formats | Structured JSON with citations, reasoning, confidence | Structured data with multiple format types |
| Citations Full | Basis framework: citations, reasoning, excerpts, calibrated confidence | Source-backed results |
| Async processing | Yes, polling-based with status updates | Yes, webhook support |
| Post-hoc enrichment | Can add enrichments to completed runs | Can add enrichment columns anytime |
| Monitoring | Separate Monitor API for ongoing tracking | Built-in Monitor feature for web tracking |
| Dashboard/UI Developer | Platform with playground | Full dashboard with chat-based enrichment creation |
| SDK support | Python SDK, TypeScript SDK, REST API, MCP Server, Vercel AI SDK tools | Python SDK, JS SDK, OpenAI-compatible endpoint, MCP Server |
## **Recall and accuracy**
FindAll has no current benchmark on parallel.ai/benchmarks[parallel.ai/benchmarks], so the evidence for its recall is how it works. You set match conditions, and FindAll only returns a candidate as matched after verifying it against every condition, with citations for each match. You trade cost for depth by choosing a generator: preview ($0.10 fixed), base ($0.25 plus $0.03 per match), core ($2 plus $0.15 per match), or pro ($10 plus $1 per match). Higher generators search more widely, so they suit queries where a missed entity is expensive.
Exa does not publish comparable recall figures for Websets. The one independent benchmark we know of, a March 2026 evaluation by NewsCatcher (32 queries focused on real-world event discovery), favored Exa: Websets reached 19.6% recall with 83.7% precision, while Parallel FindAll Core reached 5.5% recall with 77.7% precision. That's a single query set, so results may differ on other workloads.
Neither vendor has a current published recall benchmark for these products, and the only independent result favors Websets. When evaluating either product, run tests on queries that match your actual use case.
## **Developer experience**
### **Parallel**
Parallel offers a REST API, Python SDK (parallel-web on PyPI), TypeScript SDK (parallel-web on npm), and Vercel AI SDK tools, with detailed documentation at docs.parallel.ai[docs.parallel.ai]. The Developer Platform includes a playground for testing queries. Parallel also integrates with Cloudflare AI Gateway, enabling observability, caching, and rate limiting at the infrastructure level. An MCP Server is available for agent integrations.
1234567891011121314# Create a FindAll run
curl -X POST "https://api.parallel.ai/v1beta/findall/runs" \
-H "x-api-key: $PARALLEL_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"objective": "Find all AI companies that raised Series A in the last 3 months",
"entity_type": "companies",
"generator": "core",
"match_conditions": [
{"name": "ai_company", "description": "Company builds AI products or services."},
{"name": "recent_series_a", "description": "Raised a Series A round in the last 3 months."}
],
"match_limit": 100
}'``` # Create a FindAll runcurl -X POST "https://api.parallel.ai/v1beta/findall/runs" \ -H "x-api-key: $PARALLEL_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "objective": "Find all AI companies that raised Series A in the last 3 months", "entity_type": "companies", "generator": "core", "match_conditions": [ {"name": "ai_company", "description": "Company builds AI products or services."}, {"name": "recent_series_a", "description": "Raised a Series A round in the last 3 months."} ], "match_limit": 100 }'``` ### **Exa**
Exa provides SDKs for both Python and JavaScript, plus OpenAI-compatible endpoints and an MCP Server. The Websets product includes a full dashboard UI with a chat interface for creating and managing enrichments, a more accessible entry point for non-developers. Exa lists 30+ integrations including LangChain, CrewAI, LlamaIndex, Vercel AI SDK, Google Sheets, and many others.
12345678910111213# Create a Webset enrichment
from exa_py import Exa
exa = Exa('YOUR_EXA_API_KEY')
enrichment = exa.websets.enrichments.create('webset_id', params={
'description': 'Company revenue information',
'format': 'text'
})``` # Create a Webset enrichment from exa_py import Exa exa = Exa('YOUR_EXA_API_KEY') enrichment = exa.websets.enrichments.create('webset_id', params={ 'description': 'Company revenue information', 'format': 'text' })``` Parallel leans more heavily toward the developer/API-first experience, while Exa's dashboard and chat-based enrichment creation suit business users. Both now offer comparable SDK coverage.
## **Use case fit**
### **Choose Parallel FindAll when:**
**You need verified matches for exhaustive entity discovery.** Every FindAll match is checked against all of your conditions with multi-hop evaluation and returned with citations, and the pro generator spends more per candidate for hard queries where a missing entity has real cost: competitive intelligence, market mapping, compliance screening.
**You want pay-as-you-go pricing.** No monthly commitment means you can run one-off research queries without subscription overhead.
**You need deep, multi-hop reasoning for match criteria.** FindAll's explicit match evaluation stage with multi-hop reasoning across sources is purpose-built for complex boolean criteria ("Has raised Series A AND is in healthcare AND was founded after 2020").
**You're building programmatic pipelines.** The API-first design, Cloudflare AI Gateway integration, and separation of match conditions from enrichments make it well-suited for automated data pipelines.
### **Choose Exa Websets + Enrichments when:**
**You want a dashboard experience.** Exa's Websets UI with chat-based enrichment creation is more accessible for non-technical users or quick ad-hoc research.
**You need contact information enrichment.** Exa has first-class support for email and phone number extraction as enrichment types, priced at 5 credits per contact.
**You're already in the Exa ecosystem.** If you use Exa's Search API or other products, Websets is a natural extension that shares the same account, credits, and SDKs.
**You need the broadest integration support.** Exa offers 30+ integrations including LangChain, CrewAI, LlamaIndex, Vercel AI SDK, Google Sheets, and OpenAI-compatible endpoints.
**You have predictable, recurring enrichment needs.** The subscription model with monthly credits works well when you have a steady cadence of enrichment work.
## **Head-to-head summary**
| Dimension | Parallel FindAll | Exa Webset |
|---|---|---|
| Core strength | Entity discovery with explicit match evaluation | Embeddings-powered search with enrichment |
| Independent recall (NewsCatcher, March 2026) | 5.5% (Core) | 19.6% |
| Pricing model | Pay-per-query + per-match | Monthly subscription + credits |
| Lowest entry point | $0.10 (preview) | Free (1,000 credits), $49/mo (Core) |
| SDK support | Python, TypeScript, MCP Server | Python, JS, OpenAI-compatible, MCP Server |
| Integrations | Cloudflare AI Gateway, Vercel AI SDK | 30+ including LangChain, CrewAI, LlamaIndex, Vercel AI SDK |
| Contact enrichment | Via Task API | Native (email, phone) |
| Match evaluation | Explicit multi-hop reasoning | Criteria-based verification |
| Citation quality | Basis framework (citations, reasoning, confidence) | Source-backed |
## **Conclusion**
Parallel FindAll and Exa Websets solve overlapping problems with distinct approaches. FindAll is purpose-built for entity discovery at scale with a rigorous match evaluation pipeline and citations on every match. Exa Websets brings a more accessible, dashboard-driven experience with rich SDK support and the broadest ecosystem of integrations.
For teams building automated data pipelines that need explicit, auditable match criteria and per-match citations, Parallel FindAll fits well, though the only independent recall benchmark (NewsCatcher, event discovery) favored Websets, so test on your own queries. For teams that want an all-in-one platform with a polished UI and flexible enrichment types, Exa Websets is the more complete package, and business users can operate it directly. Both products are actively evolving.
By Parallel
September 26, 2026