Entity Enricher offers three enrichment strategies that control how LLM calls are orchestrated. Choosing the right strategy affects accuracy, speed, and cost — and you usually do not choose: the default is automatic selection from the shape of your schema.
From the simplest to the most powerful configuration. Each builds on the previous one.
One model, one call. The entire schema is sent in a single prompt. Fast and cheap — ideal for simple schemas or quick iteration.
Aspirin
Any entity: company, drug, legal case, research paper...
Use any LLM provider with your own API key.
Full schema in one call — auto-retries on validation failure.
Aspirin
Same strategy, but run across multiple models in parallel. Results are compared and arbitrated field-by-field to produce a single high-confidence output.
Aspirin
Any entity: company, drug, legal case, research paper...
Match — Pharmaceutical Compound
Catches type mismatches before wasting LLM credits.
Bring your own API keys — works with any LLM provider.
Full schema in one call — auto-retries on validation failure.
Aspirin
ArbitratedReasoned field-level conflict resolution produces the final trusted result.
The schema is split by expertise domain. Each model receives focused sub-prompts for each domain. Results are deep-merged per model, then arbitrated across models. Maximum accuracy for complex, multi-domain schemas.
Aspirin
Any entity: company, drug, legal case, research paper...
Match — Pharmaceutical Compound
Catches type mismatches before wasting LLM credits.
Bring your own API keys — works with any LLM provider.
Schema split by domain — self-correcting prompts retry on validation failure.
Deep merge of expertise responses per model.
Aspirin
ArbitratedReasoned field-level conflict resolution produces the final trusted result.
Leaving the strategy unset — the default everywhere, including the API and the connectors — resolves it from the shape of your schema before anything runs. The rule is deterministic, so the same schema always gets the same strategy, and the run reports which one it picked.
Expert domains needs tool calls. The strategy works by declaring one submit function per expertise domain, so a model without tool-call support is rejected with an explicit error before any tokens are spent, rather than silently degrading.
| Aspect | Single Pass | Expert Domains | Multi-Expertise |
|---|---|---|---|
| LLM Calls | 1 per model | 1 conversation per model (at least 2 requests) | N per model (1 per expertise domain) |
| Schema Sent | Full schema in one prompt | Full schema in the prompt, plus one submit tool per domain whose arguments are that domain's subset | Subset per expertise domain |
| Execution | Sequential (one call) | One generation, tool calls handled as they arrive | Parallel (all expertises run simultaneously) |
| Accuracy | Good for simple schemas | Better than single pass — each domain is validated on its own subset | Higher — focused prompts yield better results |
| Speed | Faster for small schemas | Close to single pass | Parallel execution can be faster for large schemas |
| Cost | Lower (single call overhead) | About the same as single pass — the schema is sent once | Higher (multiple calls with per-call overhead) |
| Streaming | All-or-nothing result | Progressive — one event per submitted domain | Progressive — results stream as each expertise completes |
| Partial Success | No — entire call succeeds or fails | Partly — domains submitted before a failure are kept | Yes — successful expertises are preserved, failed ones can be retried |
The multi-expertise strategy follows a four-step process for each model:
The schema is traversed recursively. Each property with an expertise domain tag is grouped with others sharing the same domain. For example, revenue and market_cap go to the “financial” group, while gmp_status goes to “regulatory”.
Each expertise group becomes a minimal sub-schema containing only its properties. This means the LLM receives a smaller, more focused prompt and only needs to fill in fields it specializes in.
All expertise calls run concurrently. A schema with 5 expertise domains will launch 5 LLM calls at the same time. As each one completes, its results are deep-merged into the accumulated output and streamed to the UI in real-time.
If some expertises fail, the system returns the merged output from successful ones with a “Partial” status. You can retry only the failed expertises, and the new results will be merged into the existing output without repeating the work that already succeeded.
Both strategies can be combined with multi-model enrichment. When you select multiple models, each model runs the chosen strategy independently. The results can then be merged using multi-model fusion to produce a single high-confidence output.