Can RAG run without sending data to a public model API?

Yes. Models can run in your own cloud account through a managed service, or an open model can run on your infrastructure. Self-hosted models need more engineering and may answer less well on hard questions, so we test the options on your data.

How do you stop the model making things up?

The model is told to answer only from retrieved passages, each answer carries citations, and we score faithfulness on a test set of your real questions. Errors drop sharply but not to zero, so citations let people check anything important.

Which model should we use?

There is no single best model. We shortlist a few by hosting rules, budget and language needs, then compare them on your questions. The integration layer is provider-neutral, so changing later is a configuration change.

Will people only see answers from documents they may open?

If the permissions exist in the source system, yes. We carry them into the index and filter at query time. Where source permissions are inconsistent, retrieval will expose that, so we flag it early.

Do we need fine-tuning, or is RAG enough?

For answering from changing documents, RAG is usually the right tool because the index updates without retraining. Fine-tuning helps with a consistent format, tone or vocabulary. We start with RAG and fine-tune only if testing shows a gap.

Language models on your systems, answering from your documents

Cited answers from your own content, with access rules and retrieval quality designed in.

See how a grounded agent is applied to inbound leads in how applyd.ai handles inbound leads.

What we build for integration and RAG

We start with the documents and questions that matter most, measure retrieval on them, and widen once the results hold.

Model calls in existing systems
Classification, extraction and drafting inside the CRM, help desk or internal tools your team already uses, with timeouts, retries and fallbacks.
RAG over your documents
Question answering grounded in your files, kept in sync as documents change, with links to the exact source passage.
Chunking and hybrid search
Retrieval tuned to how your documents are written: structure-aware splitting, keyword plus vector search and reranking.
Document-level access control
Users only get answers from documents they may open. Permissions travel with each passage and are enforced at query time.
Model selection and routing
The right model per request behind one internal interface, so a provider can change without rewriting the application.
Private deployment and retrieval evaluation
Models hosted where your data rules require, and test suites that track retrieval accuracy and answer faithfulness over time.

How a RAG pipeline works end to end

Retrieval-augmented generation has an indexing half that keeps documents searchable and an answering half that finds the right passages and writes a cited answer. A third loop measures it.

Indexing

Runs in the background whenever a source changes

Answering

Runs for every question

Not supported: search again or hand over to a person.

Evaluate and improve

On every change and in production

Step 1 of 15System work

Connect sources

Connectors pull documents, wiki pages, tickets and records, keeping each item's access list.

In practice: Incremental sync with permissions carried as metadata
You get: A source inventory with owners

Ask a question of your documents

Run a question through search, reranking and a cited answer, then switch to model only to see what a model says without the documents.

Question

How far ahead must I book flights?Can I claim a taxi home after a late meeting?Who approves leave longer than two weeks?

Answer from

With your documentsModel only

Search

Keyword and vector search return candidate passages from the example handbook.

Travel policy, section 2Book flights at least 14 days before departure where possible. Later bookings need manager approval.
Expenses policy, section 4Meals while travelling are reimbursed up to the daily limit set by finance.
Travel policy, section 5Hotels must be booked through the approved portal.
Leave policy, section 1Annual leave requests go to your manager through the HR system.

Illustrative demo. Scripted, not a live system.

How we keep answers trustworthy

Retrieval makes answers checkable. It does not make them automatically right, so we test it.

Answer only from retrieved passagesGroundedThe system cites each source and says so when your documents do not cover the question.
Permission-aware retrievalEnforcedAccess rules from the source system are applied at query time, not after the fact.
Retrieval and faithfulness scoresTestedEvery change is scored on a question set built with your experts.
Private hosting where requiredYour callYour data rules decide where models and indexes run. We test the options on your content.

Tools we commonly work with

We choose components for your data, hosting rules and budget. Naming a tool is not a partnership or an endorsement.

Model providers: OpenAI GPT
Anthropic Claude
Google Gemini
Llama
Mistral
Cohere
Private hosting: Azure OpenAI
AWS Bedrock
Vertex AI
vLLM
Ollama
Retrieval and vectors: pgvector
Pinecone
Weaviate
Qdrant
OpenSearch
Pipelines: LlamaIndex
LangChain
Unstructured
Apache Tika
Evaluation and monitoring: Ragas
Promptfoo
Langfuse
Arize Phoenix
OpenTelemetry

Common questions about this work

Tell us what you want built. We will scope it with you.

Start an AI project