Customer support
Agents get answers from manuals, product sheets and past tickets while the customer is on the line.
“Does the X200 support 220 V, and which adapter ships with it?”
RAG knowledge assistants
Ask your documents.
Check every answer.
Retrieval-augmented generation (RAG) lets an AI assistant answer from your own contracts, manuals, policies and tickets, and show the passage each answer comes from.
- Your documentsPDFs, Office files, email, SharePoint, Drive
- Parse & splitTables, scans and slides read correctly
- IndexSearch by meaning and by keyword
- RetrieveBest passages found and ranked
- Language modelAnswers only from those passages
- Answer + sourceEvery claim linked to its document
Why RAG
A general AI model
doesn’t know your business.
It has never seen your price list, your contracts or last week’s policy change, and when it doesn’t know, it guesses. RAG fixes that. Before answering, the assistant searches your documents, reads the relevant passages and answers from them.
Always current
Add or change a document and it is searchable within minutes. No retraining.
Verifiable
Every answer links to the passage it came from, so people can check it in one click.
Private
The whole pipeline, model included, can run on your own servers.
Use cases
Wherever people
search for answers.
Typical questions by team. These are examples, not client results.
Sales
Specs, pricing rules and approved answers for RFPs and security questionnaires.
“What discount can I offer on a three-year contract?”
HR & policies
Employees find answers in handbooks and benefit documents instead of emailing HR.
“How many days of parental leave do I get in Texas?”
Contracts
Find clauses and compare terms across hundreds of agreements.
“Which suppliers can terminate with less than 60 days’ notice?”
Operations
Procedures, maintenance manuals and safety instructions, searchable from the shop floor.
“What is the lockout procedure for line 3?”
IT helpdesk
Troubleshooting from your internal knowledge base, with links to the right article.
“How do I reset VPN access for a new laptop?”
How it works
The details that make answers right.
Most RAG projects that disappoint fail on parsing, search or permissions, not on the model. That is where we spend the effort.
-
Connect your sources
SharePoint, Google Drive, file shares, Confluence, email archives or databases, kept in sync so new and changed documents are picked up automatically.
-
Read documents properly
Tables stay tables, scanned pages go through OCR, slides and spreadsheets keep their structure. Answer quality starts here.
-
Index for meaning and keywords
Hybrid search combines vector embeddings with keyword search, so the assistant finds “payment terms” and also the exact clause that says “net 45”.
-
Retrieve, rerank, answer
The best passages are reranked, the model answers only from them, cites each source, and says so when the documents don’t contain the answer.
-
Respect permissions
Search is filtered by each user’s access rights, so people only get answers from documents they are allowed to open.
-
Measure before launch
We build a test set from real questions your team asks and track three numbers before go-live and after every change: whether the right passages are found (retrieval recall), whether answers stick to them (faithfulness), and whether citations point to the right place.
What you get
A working assistant,
not a slide deck.
- The assistantIn a web app, Microsoft Teams, Slack or your intranet, wherever your team already works.
- Connectors kept in syncYour sources indexed and updated automatically.
- Permission-aware searchAnswers drawn only from documents each user can access.
- Evaluation reportAccuracy measured on your own questions, before and after launch.
- Admin viewSources, usage, unanswered questions and user feedback.
- Documentation & handoverYour team can run it, or we run it for you.
RAG or fine-tuning
Two tools.
Different jobs.
Fine-tuning changes how a model writes and behaves. RAG gives it your facts. Most projects start with RAG and add fine-tuning only if needed.
| Criterion | RAG | Fine-tuning |
|---|---|---|
| Teaches the model | Facts from your documents, looked up at question time | Style, format, vocabulary and narrow tasks |
| When content changes | Add the document; searchable in minutes | Retrain the model |
| Shows its sources | Yes, for every answer | No |
| Respects document permissions | Yes, per user | No; everything trained in is visible to everyone |
| Usual order | First | Later, if the pilot shows a need |
Where it runs
Your choice of model.
Your choice of place.
The assistant works the same way in all three setups. What changes is where your documents and questions are processed.
Cloud API
Claude, GPT or Gemini under business terms that exclude training on your data. Fastest to start, paid per use.
Private cloud
An open-source model on GPUs rented in your own cloud account. No outside AI provider involved.
Your hardware
An open-source model on a server in your office or data centre. Nothing leaves your network.
Open-source LLMsTechnology
Proven, open tools.
No lock-in.
We pick components for your volume, budget and privacy needs, and you own the result.
Models
- Llama
- Qwen
- Mistral
- Gemma
- gpt-oss
- Claude, GPT or Gemini via API
Search & storage
- pgvector
- Qdrant
- Elasticsearch
- OpenSearch
Document parsing
- Docling
- Unstructured
- Tesseract OCR
Frameworks
- LlamaIndex
- LangChain
- Haystack
Evaluation
- Ragas
- promptfoo
- Your own test questions
When RAG isn’t
the right tool.
- Your team asks the same few questions. A well-organised FAQ or intranet page is cheaper and just as fast.
- The answers live in a database, not in documents. Live figures such as stock or balances are better served by connecting the assistant to the system itself, through MCP.
- Your documents contradict each other. Fix the sources first. An assistant will faithfully repeat the confusion.
How we work together
Start small.
Decide on facts.
Every project goes through the same three stages. You decide at the end of each one whether to continue, and you can stop after any stage and keep everything built so far.
-
Stage 1
Assessment
We look at your process, your data and your systems, then tell you plainly whether AI is worth it, which approach fits, and what it would cost to build and run.
You get a written recommendation with scope, timeline and cost.
-
Stage 2
Pilot
A working version on your real data, used by a few of your people, with results measured against questions and cases from your own business.
You get a working pilot and an evaluation report.
-
Stage 3
Production & support
We secure it, connect it to your systems, roll it out to your team and keep it running: monitoring, model updates and improvements.
You get a production system, documentation and ongoing support.
Yours to keep
- Source code, in your own repository
- Prompts, configuration and infrastructure setup
- Test sets and evaluation reports
- Fine-tuned model weights, if any
- Documentation and runbooks
Not included in our fees
- Hardware, if you buy or rent GPUs
- Cloud hosting
- AI model usage, if you use a cloud API
- Third-party software licences
We list them in the assessment, so the full cost is clear before you commit.
Questions
Will our documents be sent to OpenAI or another AI provider?
Only if you choose to. We can run the entire pipeline, including the language model, on your servers or in your private cloud. If you prefer a cloud model for quality or cost, we use providers’ business terms that exclude training on your data.
Which file types and sources can it read?
PDF (including scanned documents, through OCR), Word, Excel, PowerPoint, email, web pages and plain text, from SharePoint, Google Drive, Confluence, file shares or databases. If your content lives somewhere else, we write a connector.
How do you stop it from making things up?
The model is instructed to answer only from the passages it retrieves, to cite them, and to say when the documents don’t contain the answer. We then measure this on a test set of real questions before launch, and keep measuring afterwards.
Does it respect who can see what?
Yes. Search results are filtered by each user’s permissions in the source system, so an employee never gets an answer drawn from a document they could not open themselves.
How long until we see something working?
A pilot runs on a sample of your real documents, with a handful of users, and ends with an evaluation report. We agree its scope and timeline in writing during the assessment.
What does it cost to run?
It depends on the number of users and documents, and on whether the model runs on your hardware or through an API. The assessment gives you a running-cost estimate for each option before you commit.
We already have a RAG assistant and its answers are poor. Can you fix it?
Often, yes. We first measure it on a test set of your questions to find where it fails. That is usually document parsing, search or how documents are split, rather than the model. Then we fix those parts and keep what works.
What do you need from us?
Access to a representative sample of documents, 20 to 50 real questions with the answers your experts would give, and one person who can judge the results.
Other AI solutions
Start with
a conversation.
No call needed to start. Send a short brief and we’ll reply in writing with what we’d do and what it would take.
Houston, Texas · Working with clients worldwide