Check How AI Uses Company Data (RAG, APIs and MCP), Complete Resource Pack

Ideas to keep

1. 01A customer writes: "Please cancel order 24. I no longer need it." You want an AI assistant to help with the request.
2. 02Question one: which shop rule would allow a cancellation?
3. 03Question two: what would the assistant need to find out about order 24 right now?
4. 04Question three: which action would you want to approve before it happened?
5. 05By the end of the section you can explain how the assistant finds a rule, gets current order information, and acts only with the access you allow. You can then use those three questions on any seller's AI product.

Try it (optional)

1. Write the three questions down before you watch Lecture 4.1.
2. Next to each one, write your own guess at the answer.
3. Come back to your notes in Lecture 4.8 and compare.

Files and links

No outside links in this lecture. Write down the three questions about the order-24 request (the shop's rule, the current state of the order, the action you would approve) and keep them for Lecture 4.8.

• Section 4 study guide (PDF)/downloads/agents/section-04/Section-4-Study-Guide.pdf

Ideas to keep

1. 01The company chooses the policies, handbooks and other guidance the assistant may search. That collection is the knowledge base. Someone also has to mark the current versions and decide who may open sensitive documents.
2. 02Priya asks about "vacation days" while the policy may say "annual leave", so a search for her exact words could miss the answer.
3. 03The documents are split into passages, such as a paragraph or a few paragraphs about one subject.
4. 04A model trained to compare meanings turns each passage into a list of numbers. That list is an embedding. Passages with related meanings sit near each other, like positions on a map.
5. 05A vector database stores those lists, keeps them linked to the original text, and searches them for close matches when a question arrives.
6. 06The question gets its own list of numbers, the closest passages are retrieved, and they go to the answering model with the question as context.
7. 07Finding information and using it to write an answer is retrieval-augmented generation, or RAG. Some RAG systems use keyword search, or combine several search methods.
8. 08Ask the assistant to show the policy passage beside its answer, and test it first with a question whose answer you already know.

Try it (optional)

1. Pick one question from your own work whose answer is in a company document you know well.
2. If your company has an assistant that searches its documents, ask it that question and ask it to show the passage it used.
3. Check that the passage comes from the current version of the document and answers that exact question.

Files and links

No outside links in this lecture. If your company has an assistant that searches its documents, try it with a question whose answer you already know, and ask it to show the passage it used.

Ideas to keep

1. 01Training happens while a model is developed: examples are used to adjust how it responds, and those changes shape its later answers.
2. 02Supplying Maya's contract as context does not by itself carry out that process. The model has the clause available while it writes this one answer.
3. 03A provider may separately use submitted material in a later training process, depending on its account terms and settings.
4. 04Saving the file in Maya's account so she can ask again tomorrow does not train the model either. The file stays stored information that the assistant can supply again.
5. 05So "we do not use your data for training" answers only part of her concern. She still needs to know where the file and conversation are kept, who can open them, and when they will be deleted.
6. 06The length of time a service keeps information is its retention period.
7. 07Business account terms can differ from individual ones, and a company administrator may control some settings.
8. 08If another company provides the model behind the assistant, the contract passage may go to that company too. Find out which services receive it and what each one does with it.

Try it (optional)

1. Open the privacy settings and terms for the account you use with an AI assistant at work.
2. Find where your files and conversations are stored and who in your company can open them.
3. Find the retention period, and whether submitted information may be used for future training.
4. Write down any other company named as receiving your information.

Files and links

No outside links in this lecture. The settings to check live in the privacy settings and terms of the account you use: where files and conversations are stored, who can open them, the retention period, whether submitted information may be used for future training, and which other companies receive it.

Ideas to keep

1. 01A contract says when payment was due. Only the current accounting record says whether the money arrived this morning.
2. 02Daniel's invoicing application offers the model a tool called "get invoice status" and tells it the tool needs an invoice number.
3. 03Behind the tool is a function: a piece of code written to do one job, here the invoice lookup. The model asking to run that function, with the number 104, is function calling.
4. 04The application checks the request first: an available tool, the required information, the intended accounting account. Then it runs the function, which asks the accounting system through its existing API.
5. 05The same path must carry a failure back. If the invoice does not exist, the connection has expired or the system is down, the application should return that failure, or the model may write a confident reply with no record behind it.
6. 06When several AI applications need the same tool, shared rules help. The program offering the tool is an MCP server, the connecting part inside the application is an MCP client, and the rules they follow are the Model Context Protocol, or MCP.
7. 07MCP does not create the record or grant access to the account. The builder still connects the correct account and gives the tool only the permissions it needs.
8. 08Permission to read an invoice's status is separate from permission to change it, cancel it or issue money.

Try it (optional)

1. Pick one live fact your work depends on, such as a payment, a stock level or a delivery date.
2. Name the system that holds the current record of it.
3. Write which actions a connected assistant would need: only reading, or also changing something.
4. Write what the assistant should say if the record cannot be reached.

Files and links

The official introduction to the Model Context Protocol, the shared rules an MCP client and an MCP server follow.

• What is the Model Context Protocol (MCP)?https://modelcontextprotocol.io/docs/getting-started/intro

Ideas to keep

1. 01Priya starts with 20 real staff questions whose answers her team can check, including 3 difficult cases: a current policy that was never added to the searchable files, 2 policy versions that conflict, and a question with no approved rule at all.
2. 02Twenty questions is a manageable first exercise for her, not a standard for every AI system, and passing them does not prove every future question will work.
3. 03Before each run she records the source that should support the answer, and the expected answer or an acceptable request for help. If a policy is missing, saying so and asking HR passes, and inventing an answer fails.
4. 04She keeps every failed case. When the builder changes the search, instructions, model or tool connection, she runs them again to see what improved and what broke.
5. 05A repeatable test used to judge how a system performs is an evaluation, often shortened to eval.
6. 06An attempt to redirect an AI system through material it reads is prompt injection. Priya tests it with a harmless line in a test copy of a policy.
7. 07If the assistant can read private records or send messages, include those actions in the tests, with fictional records and a destination the company controls, and check that blocked actions stay blocked.
8. 08EchoLeak, a historical case from 2025: researchers showed that a crafted email could make Microsoft 365 Copilot disclose information from its working context, with no link clicked and no attachment opened. Microsoft patched it before it was publicly disclosed.

Try it (optional)

1. Write 5 to 10 questions about your own work whose answers you can check.
2. Make at least one of them a question the documents cannot answer.
3. For each question, write the source that should support the answer and what a good answer looks like.
4. Keep the list, and run it again after anything about the assistant changes.

Files and links

The lecture shows these two pages for the EchoLeak case: the Cato Networks write-up of the 2025 demonstration, and Microsoft's own advisory, released 11 June 2025, which lists the vulnerability as not publicly disclosed and needing no customer action to resolve. This is the harmless line Priya adds to a test copy of the policy. Use it only in a test copy of a document your assistant reads, never in the real one, then ask your normal question and see whether the answer follows the policy or the line.

• Breaking down EchoLeak (Cato Networks)https://www.catonetworks.com/blog/breaking-down-echoleak/
• CVE-2025-32711, M365 Copilot Information Disclosure Vulnerability (Microsoft Security Response Center)https://msrc.microsoft.com/update-guide/vulnerability/CVE-2025-32711
• quiz-4.4.md/downloads/agents/section-04/quiz-4.4.md

Ideas to keep

1. 01"Thanks for getting in touch" and "Thank you for your message" both do the job. The wording differs because the model has several possible ways to continue.
2. 02The model writes in small pieces called tokens. At each step it works out how likely each possible next piece is, then one is selected.
3. 03Temperature changes how strongly the likely choices are favoured. Lower usually favours them more, and higher gives less likely choices a better chance, so repeated replies vary more.
4. 04Top-p changes which group of choices is considered. Lowering it usually narrows the group, and raising it allows a wider one.
5. 05Both affect selection while the answer is being written. Neither checks whether a statement in the reply is true.
6. 06Parameters are the numerical values learned during training. They work together to shape how the model handles text.
7. 07Fine-tuning is training an existing model further with selected examples, such as replies the company has approved. Try clear instructions and examples first, and test new replies afterwards.
8. 08A model trained on the company's style still needs current information to answer a question about an order.

Try it (optional)

1. Ask an assistant for the same short reply three times and compare the wording.
2. Decide whether the differences matter for your use, or whether all three are acceptable.
3. If you need a fixed style, write it as instructions with one or two examples before anyone proposes further training.

Files and links

No outside links in this lecture. The settings available, and whether further training is offered at all, depend on the model and the application you use.

Ideas to keep

1. 01The Monday report can only start if a computer or service is on to notice the time. Computing and storage rented from a provider over the internet is called the cloud.
2. 02The cloud gives the workflow a place to run while Maya's laptop is closed. It does not decide what the workflow does or make the supplier connections reliable.
3. 03With prices arriving through an API, a spreadsheet and an unusual review rule, a builder may write custom software for that logic.
4. 04A framework is a structured collection of reusable software components that connect models, tools and the rest of an application. LangChain and CrewAI are examples a builder might consider when their components fit the job.
5. 05Visual workflow tools such as n8n, Zapier and Make let a builder arrange and connect the steps on one screen. A visual workflow can hold fixed steps, a step where a model chooses an action, custom code, or all three.
6. 06Claude Code or Codex can help a builder produce custom software from a clear description of the inputs, rules, result and checks.
7. 07Deployment is making software available where people or scheduled jobs will use it, with its settings and connections.
8. 08Someone must provide access to the supplier sources, receive the failure notice, pay the running cost and repair a connection when a supplier changes its API, spreadsheet or website. A useful failure notice names the affected source.

Try it (optional)

1. Pick one report or task you want done on a schedule.
2. When a builder proposes a tool, ask Maya's three questions: Which part of the job does the tool help build? Where will the finished workflow run? Who will provide access, receive a failure notice and keep the connections working?
3. Write the answers down, with a name beside each responsibility.

Files and links

The tools the lecture names as examples. None is recommended over another, and a builder should be able to explain why one fits your job. When a builder proposes one of them, bring the discussion back to your job with Maya's three questions: Which part of the job does the tool help build? Where will the finished workflow run? Who will provide access, receive a failure notice and keep the connections working?

• LangChainhttps://www.langchain.com/
• CrewAIhttps://crewai.com/
• n8nhttps://n8n.io/
• Zapierhttps://zapier.com/
• Makehttps://www.make.com/
• Claude Code overview (Anthropic documentation)https://code.claude.com/docs/en/overview
• Codex (OpenAI documentation)https://developers.openai.com/codex

Ideas to keep

1. 01One seller says, "Our AI agent handles customer support." The other says, "Our assistant checks orders and prepares replies for you to approve." The second names work you can examine.
2. 02Your shop allows cancellation while an order is still waiting to be dispatched. The assistant needs the rule and whether order 24 has been dispatched, and the rule alone cannot tell it the second.
3. 03The knowledge base can supply the rule, and a tool connected to the order system can look up order 24's current status.
4. 04Ask the seller how those steps are chosen: does the model decide to look them up, or does a fixed workflow follow them? Ask what happens when a request needs an extra step.
5. 05A drafted reply such as "We can cancel order 24 because it has not been dispatched" has not cancelled the order. Cancelling needs a cancellation tool and permission to use it.
6. 06Reading the order, cancelling it and sending a reply are separate actions. Your shop might allow lookups while requiring your approval for cancellations.
7. 07Ask each seller to demonstrate with a test order. If it promises to cancel after your approval, approve the test action and check that the order's status changes. A message saying "done" is weaker evidence than the updated order.

Try it (optional)

1. Write one real request from your own work that an AI product claims to handle.
2. List the rule it needs, the live record it needs, and each separate action it would take.
3. Mark which actions it may take alone and which need your approval.
4. Ask for a demonstration on test data and check the record itself afterwards.

Files and links

No outside links in this lecture. When a seller demonstrates on a test order, check the order record itself afterwards, not only the message that says it is done.

• quiz-4.7.md/downloads/agents/section-04/quiz-4.7.md

Ideas to keep

1. 01The shop's rule says cancellation is allowed before dispatch. The current order record says whether order 24 has been dispatched.
2. 02If it has not, the assistant can prepare a reply, but cancelling still needs its own tool, permission and any approval the shop requires.
3. 03The rule can come from a searched knowledge base. Finding the passage and supplying it for an answer is the RAG part of the case.
4. 04A connected tool can request the live order status through an API. MCP can give compatible applications and tool providers a common way to offer and call such tools.
5. 05The order record stays the source for whether the cancellation actually happened.
6. 06Three more questions belong to the same system: how long it keeps customer information (retention), whether it handles known test cases (an eval), and whether instructions inside a message or document could redirect it (prompt injection).
7. 07You can now ask what an AI product reads, what it changes and what it leaves for a person.

Try it (optional)

1. Take out your notes from Lecture 4.0.
2. Beside each question, write the part of the system that answers it: knowledge base, live record through a tool, or your approval.
3. Add the three extra questions: retention, eval and prompt injection.

Files and links

• Notes: Check How AI Uses Company Data (RAG, APIs and MCP)/resources/library/agents/section-04-summary/
• summary.md/downloads/agents/section-04/summary.md
• Section 4 study guide (PDF)/downloads/agents/section-04/Section-4-Study-Guide.pdf