Bring your own key. Keep your own taste.
We pick
your side.
OpenAI. Anthropic. Google. xAI. Open models. Choose from supported models and providers through Salsa’s BYOK connections and model catalog. Your work stays in Salsa when your model choice changes.
OpenAI Anthropic Google xAI Open models
A stronger model can bring stronger reasoning. Salsa supplies the context, memory, tools, and execution around it. Choose for the task, your budget, and your data requirements—without rebuilding your business around a model vendor.
Available models, modalities, and tools depend on the configured provider. BYOK means API credentials; consumer chat subscriptions do not automatically cover API usage. Results depend on the model, evidence, tools, and workflow together.
A recipe with different jobs.
Retrieve facts.
Teach behavior.
Let agents act.
Retrieval supplies current policies, case law, product knowledge, and business records. Updating a source should not require retraining a model.
Fine-tuning can teach consistent extraction, formats, domain language, and tool-use patterns using curated, authorized examples. Tune an embedding model or reranker separately if retrieval evaluation identifies that bottleneck.
Agents compose the pieces: select sources, retrieve evidence, reason over it, call allowed tools, check the result, and preserve useful state. A fine-tuned model can sit inside the same loop as RAG.
Start with a prompt-and-retrieval baseline. Train only where held-out evaluations justify it. Keep training examples separate from the final test set, and verify that the provider supports the chosen tuning method and deployment.
Method: OpenAI’s evaluation, prompting, and fine-tuning workflow. Training a domain model is a configured project, not an automatic consequence of using Salsa.
Same kitchen. Your specialty.
Make it fluent
in your business.
Three example deployments. Each begins with your sources and ends with work you can inspect.
01 / LEGALFind the clause.
Follow the authority.
Datasets: authorized contracts and amendments, clause playbooks, matter files, current statutes, and relevant case-law opinions. CourtListener is one source to evaluate; licensed research collections require their own access.
What a legal workflow needs
Contract clauses and case-law passages need different indexing. Preserve section boundaries, defined terms, cross-references, jurisdiction, court, decision date, and document version. Retrieve adverse authority as well as supporting cases.
For a clause review, compare every required clause type against the complete document and expert labels. For research, check citation accuracy, authority, and current treatment through appropriate research services and qualified legal review. Retrieval alone cannot certify that a case remains good law.
Deliverable: a review table with passage-level citations, omissions, conflicting evidence, and questions for counsel.
Know the answer.
Know when to hand off.
Datasets: approved help articles, product versions, policies, resolved support conversations, and authorized account records. Exclude private notes and stale answers that customers should never see.
What a support workflow needs
Filter by account permissions, product, locale, and effective date. Retrieve the policy, use tools for live order or subscription state, then draft a grounded response or escalate with context.
Test factual correctness, unsupported promises, sensitive-data exposure, escalation accuracy, and whether the customer’s problem was actually resolved. Reopen rates and human review help catch apparent successes that were not useful.
Deliverable: an answer or reviewable action with its source and a useful handoff. Customer-facing deployment needs configured channels and validated procedures.
Trace the loss.
Explain the dollars.
Datasets: permissioned authorization, settlement, refund, chargeback, dispute, and ledger records; policy versions; and eventual loss outcomes. Define whether “loss” means fraud, fees, unrecovered disputes, or reconciliation breaks.
What a transaction-loss workflow needs
Compute totals and reconcile records with deterministic queries. Use risk models for calibrated predictions; use RAG for policies and prior investigations. Keep the agent’s explanation tied to the underlying transactions.
Evaluate with time-based splits and entity separation where appropriate. Account for delayed chargeback labels and information unavailable at decision time. Measure precision, recall, false-positive cost, dollar-weighted loss, and reconciliation accuracy.
Deliverable: an evidence-backed investigation and proposed next step. Financial actions need explicit authorization; neither retrieval nor fine-tuning guarantees recovered funds.
The composition framework.
The intelligence needs
somewhere to work.
Load context→Plan→Use tools→Inspect→Record & continue ↺
Boot with the right context
Establish identity, team scope, permissions, model configuration, available tools, and budget. Load the relevant workspace memory and run state before an agent begins.
Give each ingredient a contract
Compose retrievers, rerankers, models, specialist agents, and tools through explicit inputs and outputs. Version the pieces so an improvement can be evaluated and a failure traced.
Make the loop accountable
Salsa connects event-backed state, tool execution, and plan review. A production domain harness also needs bounded retries, idempotent actions, checkpoints, and clear stop conditions. Review consequential actions before execution.
Meet the work in its own format
Documents, tables, images, and voice need different handling. Preserve the original and its provenance through parsing, OCR, transcription, or structured queries. STT lets you direct work by voice; TTS brings the response back to you.
Salsa supplies the shared workspace, memory, agent, model, and execution foundation. Domain datasets, licensed content, learned rerankers, fine-tuned models, and domain-specific production guarantees require configuration and validation. “Enterprise-class” describes the ambition and toolset; it does not imply an unlisted certification or service-level agreement.
See Salsa’s execution and approval flow ↗
