Everything that happens behind your agent, explained simply

Step by step

1. 1

   We take the document

   We read it with its headings and tables. If it is a scan or a photo, we use OCR to recognise the text. Diagrams are described in words.

2. 2

   We clean it

   We remove the headers repeated on every sheet and the page numbers.

3. 3

   We split it by topic

   We cut where the meaning changes, not every so many words. That way each piece makes sense on its own. Each one is labelled with the file and section it came from, plus a line that places it.

4. 4

   We generate vectors

   Each piece becomes a list of numbers that represents what it means.

5. 5

   We store them

   The vectors go to a vector database. From there, the agent finds them by meaning.

• Spreadsheets are split by table

  Each piece repeats the header row, so "price" always stays attached to its column.

• Replacing is safe

  The previous document is deleted only when the new one is ready.

• Each agent has its own

  An agent only finds the documents uploaded for it.

1. 1

   The channel notifies us

   We use Meta's official WhatsApp API. When someone writes, Meta notifies the platform through a webhook, and the platform checks the notice is authentic before doing anything. Instagram works the same way.

2. 2

   The message enters a queue

   Before it is processed it is stored in a message queue. That way it is not lost if things get busy or something fails in the middle: it is retried.

3. 3

   Messages in a row are grouped

   People write in pieces. The messages one person sends in a row are grouped and answered together, with all the context.

4. 4

   It checks everything is in order

   That the agent is active, that the plan has answers left and that nobody on the team has taken over the conversation.

5. 5

   The agent prepares the answer

   It reviews the conversation, searches your documents if needed and writes. If it has to do something (book, record, notify), it does it at this step.

6. 6

   The answer returns through the same channel

   It goes out the way the message came in, through the same official API, and everything that happened in between is recorded.

• Voice notes, images and ads

  Voice notes are turned into text before reaching the agent, and it receives images too. If the person arrives from an ad, the agent knows it, so you never lose where they came from.

• A person can take control

  From the panel, or with a keyword in the same chat, your team takes over the conversation. Meanwhile the agent does not answer.

• Your website chat is more direct

  There is no webhook or queue: the chat asks the platform for the answer and gets it right away, because the person writing is looking at the screen.

The technology underneath

• Vector database

  Each piece of your documents is stored as a vector: a numeric representation of its meaning. Searching compares meanings, not letters, so it finds "how much is it?" even if the document says "service fee". We use pgvector on Postgres, and Pinecone for agents with a lot of material.

• Semantic chunking

  Documents are cut where the topic changes, not every so many words. Each piece makes sense on its own and carries the file and section it came from, so the search brings back whole ideas, not loose scraps.

• Context per chunk

  Before it is stored, an AI writes each piece one or two lines that place it inside the document. An "expires June 30" stops being an orphan: you know which contract it is about.

• Hybrid search

  By meaning it finds texts that talk about the same thing in other words, but it mixes up codes, prices and proper names. By exact word it does the opposite. Both are used at once and the results are merged.

• Re-ranking

  Instead of keeping the first results, many candidates are fetched and a re-ranking model rereads them one by one against the question and orders them by how well they answer it. It is the difference between "something similar" and "the fact".

• A well-formed question

  People write in a hurry: "how much the thing 2 lts". Before searching, a light AI completes the question with the industry's words, and the agent still sees what the person wrote.

What your business gains

• Proven techniques

  Context per chunk, hybrid search and re-ranking are the combination Anthropic published as Contextual Retrieval, and the one we use, adapted to Deployate.

• The agent knows which documents it has

  Before searching it sees the list of what you uploaded, with its summary. Knowing there is a "2025 price list" lets it decide what to search and with which words.

• Less filler, better memory

  The more text an AI receives, the worse it remembers what matters. That is why it gets just enough: the trimmed conversation and only the pieces useful for that question.

• If one piece fails, there is still an answer

  If the AI that completes the question or the re-ranking does not respond, the search carries on the simple way. It never stalls the conversation.

• Measured, not guessed

  Each search improvement is tested with the same set of questions, before and after, to compare how many found their source.

Two ways of searching

A turn from the inside

1. 1

   The message arrives

   Your agent starts with the conversation, its instructions and the map of your documents: what exists and in which section. We do not hand it the content in advance.

2. 2

   It searches or reads

   It asks for what it needs: it searches your documents or reads a whole section. The one who knows what is missing is the one who read the conversation, not a search run with the message as it came.

3. 3

   It judges what it found

   Does it answer the question? If not, it searches another way. It does not accept a search just because it brought something back.

4. 4

   It acts if needed

   Books, records, sends an email. Every action that changes something is recorded before it is done, and if it cannot be recorded, it is not done.

5. 5

   It answers

   When it has what it needs it replies, in your business's tone, and the turn ends.

What changes for your business

• The one who answers is the one who judges

  The same agent that is going to reply reads the search result and decides whether it is good enough. There is no separate step accepting a search that brought something but did not answer.

• It corrects course on every pass

  It can change its mind after each step: rephrase the search, read more, or decide it already has what it needs.

• It does not work more than needed

  If the first reading already has what is needed, it answers. It does not keep searching just in case.

• It does not invent what is missing

  If to do something it needs a fact only your customer can give, it asks. It does not fill the gap with a guess.

Who decides the turn is over

• Looking is not logged

  Searching and reading your documents changes nothing, so repeating it does no harm. Only what has an effect is logged: booking, sending, recording, building a document.

• If something was left halfway, it says so

  If the conversation was cut right while a tool was working, it is not repeated. Your agent tells your customer it could not confirm whether it got done.

• Minutes-long tasks do not block

  A long job starts, your customer gets a notice that it is being prepared and the conversation is free.

• When it ends, the agent picks its turn back up

  When the long job finishes, your agent continues from where it left off, with its tools and with what it already did in view: it can build the document, send the email or start another job. There is a cap so it does not chain forever.

• Several steps in a row, with a cap

  The agent can chain actions, but to keep costs in check there is a limit of 3 rounds and 6 actions per message. If it asks for exactly the same thing twice, the second one is not run (unless the first one failed, in which case it is retried).

• Automatic tags

  After each answer, the conversation is tagged with its intent ("price inquiry", "wants to book a visit"). It helps sort contacts and see what you're asked most.

• Messages to start a conversation

  From the panel you manage the templates approved by WhatsApp, which are the messages you can use to write to a customer first.

• No connection, no action

  A tool you haven't connected yet simply isn't used. The agent doesn't pretend to have done something it didn't do.

• A timeline per answer

  What the customer wrote, each step the agent took and the final answer. On top, a summary: how many searches, how many actions, how long it took and what it cost.

• What it searched for and what came back

  For each search you see the words the agent used, how many chunks it found and which document each one came from. If it found nothing, that's flagged.

• Why it's worth a look

  A search that keeps coming back empty says a document is missing. It's the difference between guessing why it answered that way and seeing it.

Two grades, not one

• The search grade

  Did the agent find the chunk of the document where the answer is? We check whether it arrived, where it ranked, or whether it was among the candidates and got left out.

• The answer grade

  Another AI acts as the judge: it compares what the agent answered with what was expected and gives a score, a verdict and the reason. It doesn't compare word by word: saying the same thing in other words is fine.

How to read an exam

Four difficulty levels

Tools take the exam too

• They never run for real

  During an exam no email is sent, nothing is booked, nothing is charged and nothing is recorded in your customer system. Tools reply with a simulated result.

• Two cases per tool

  One where the tool works and one where it fails. In the second, the agent has to say it couldn't be done: it can never tell your customer something was done if it wasn't.

• Without using the tool, it doesn't pass

  Whether the agent used it is recorded. If it didn't, the case isn't accepted no matter how polished the answer is.

• What isn't tested yet

  Only what can be asked in a single message is evaluated, and we check that the agent uses the tool, not which data it uses it with. Both are pending.

What we use it for

• Before and after every change

  Each exam saves the configuration it was taken with. When we change something in the search, we run the same exam again and compare: measured, not guessed.

• The grade per level says more than the average

  An agent scoring 90 on basic questions and 40 on complex ones doesn't answer "so-so": it struggles to cross information. The average alone would hide that.

• A constant health check

  Each agent can have a control question with its expected answer. It's asked automatically, graded by the same judge, and it only passes with a "correct" verdict.

• A waiting room for every message

  Each message stays stored until it's answered. If something goes down in between, it's retried automatically.

• Backup from another provider

  If the main AI provider goes down, the answer continues through a different one, with the same instructions and the same information. Your customer doesn't notice.

• Actions aren't repeated

  If an answer has to be retried, what the agent already did (booking a visit, sending an email) isn't done again.

• A failing action doesn't cut the conversation

  If a tool fails, the agent learns the action was not done and tells your customer. They are neither left without an answer nor shown technical errors.

• Review by a person

  A message that failed every attempt isn't discarded: it's set aside for the Deployate team to review. It happens very rarely, and almost always because of a long provider outage.