In plain words
What stands around the model
Where the conversation happens
Channels- WebsiteMobile app
- Messengers
- Telegram
- Next channel
The harness: one pipeline for every channel
Orchestrator- Saves every message first
- Queues one reply per conversation
- Builds the turn: instructions, tools, history
- Records every step
The core, left as it is
Custom for each business: configured, or coded
Deterministic workflow- Configured
- Fields to collect
- Outcomes and their criteria
- Consent before contact
- Handover to a person
- Coded
- An app of its own, such as ads analytics
Next stepA model per assistant, replaceable
AI- Claude
- Claude in the EU
- GPT
- Fallback model
Next modelWhat the model can reach
Tools and knowledge- Knowledge
- Search
- Whole documents
- Tables
- Tools
- Calendar
- CRM
- Price calculator
- Connected
- Your own system, over MCP
Next system- Closed GDPR perimeterEU mode, set per client: website, mobile app, hosting, database, models and embeddings stay inside the EU. A processing agreement, export and deletion for every client.
- Where do the answers come from?
- From the company's documents and tables. Each demo page shows them next to the chat, word for word, so an answer can be checked against its source.
- Can it use our own systems?
- Yes, live. What a business needs is configured, coded, or connected: a system that speaks MCP, such as an ERP or a shop, brings its own tools, and the assistant asks it while it answers. See Connected over MCP and the parts & orders demo.
- What if it sees only a part?
- It says so. An assistant never answers from a part of the data without saying so. A list cut short is called incomplete, and a document that did not arrive whole is not judged.
- How do I know it works?
- Scripted conversations are run against a copy of the assistant, and every report is read in full. Each turn leaves a record: what it looked up, what it did, what it cost.
- What does it cost to run?
- One to two cents for a simple reply on the model I use, two to four for a full offer check or a CV scorecard.
Under the hood
Data is prepared before it is searched
Each assistant is a folder in a repository: instructions, settings, documents, tables. One script uploads it, a second checks that the running assistant matches it. Nothing is edited by hand on the platform.
Documents are cut into passages of about 300 tokens, along their structure:
- A heading stays with the text under it.
- A long table is cut between rows, and every part repeats the header row.
- A paragraph is cut between sentences, never inside a word.
- Every passage carries the title of its document and the headings above it. Both take part in the search.
Against a cut by character count (1,200 characters, overlap 150), on the 80 documents of the seven assistants on this site:
| Passages that … | Cut by characters | Cut by structure |
|---|---|---|
| exist in total | 166 | 213 |
| begin in the middle of a word | 59 | 0 |
| end with a heading whose text went to the next passage | 41 | 0 |
| do not know which document they come from | 86 | 0 |
Facts that are counted or filtered are also held as tables. A script builds each table from its document, so the two cannot disagree. A cell holds only what the document states: no word about a lift is an empty cell, and empty is not "no".
One PostgreSQL database holds all of it: documents, passages with their embeddings, table rows, conversations and the trace of every turn. There is no separate vector store: with 14 to 76 passages per assistant a comparison with every passage is fast, and the vectors stay under the same access rules and the same deletion as their text.
Search is a loop the model runs
The calls in the figure come from test runs of the demos.
A question in free words
Search
search_knowledge("Error 204 C axis does not reach target position")
- Ranked by meaning, with embeddings
- Ranked by words: a rare word weighs more
- The two orders are merged
A document that has to be read whole
Document by code
get_document("RQ-220")get_document("EP-01")get_document("OF-A")
- The model sees a catalogue: code, title, size
- It takes several documents in one step
Anything counted or filtered
Exact query
query_data(table: listings_sale, where: rooms >= 3, price_eur <= 500000, about: "needs renovation")
- The model names a table, columns and conditions, never SQL
- The platform builds the query from a checked schema
- It runs under a role that reads only this assistant’s rows
Table "listings_sale": 4 rows match; all 4 are shown, ordered by closeness to "needs renovation".A wrong call comes back with the valid names, and the model corrects it.- Search does not assemble a document. The query "check offer A against the requirements" returned no passage of offer A among the best ten. So the offer check and the CV scoring take their documents whole, by code.
- The model never writes SQL. The pattern is the one from What a semantic view really does.
- Meaning sits on top of the exact filter. For "three rooms or more, up to €500,000, needs renovation" in the estate agency demo the conditions decide which listings fit: exactly four, the same four every time. The wish in free words only puts the closest one first.
Tests start below the model
The free checks come first and cost nothing to repeat. A run with the model is paid and starts with an estimate.
Instructions, data or settings, in the folder of the assistant.
No model · free
The running assistant is compared with its folder in the repository. The tables are rebuilt from the documents.
No model · free
Seventeen statements about the data, asked the way a tool asks, without the model.
With the model · cents a turn
A script plays the visitor through the chat interface, then reads the answers, the events and the trace.
- Scenarios
- Grounding
- Red team
- Repeats
A person · reads
The report of the run is read in full: every dialogue, the cost, a written assessment.
A short check after rollout.
- Reference results make scoring testable. The four made-up candidates of the CV demo have to score 95, 73, 68 and 14. The three offers have to come out as 11, 7 and 6 compliant requirements out of twelve, with the exact split between deviation, unclear and not addressed.
- Checks catch the mechanics, reading finds the defects. The checks are patterns over the answer and the trace. One CV scored 68 in one answer and 73 in another, and none of them failed. The fix went into the data: the vacancy now defines each rating for each requirement.
- Staging, not live. On the live assistants a test run would fill the working inbox, open cards in the CRM, take calendar slots and use up the daily reply limit.
- Cost is measured. Seven cases with eleven visitor turns: all green, 26 cents. An offer check that searched seven times cost 12 to 16 cents; with three whole documents in one step, two to four.
See it work: offer check, CV scoring, estate agency, parts & orders on a live ERP, or all demos. Each page shows the data next to the chat.
The platform under these demos is my own. It runs live, and it is not open to external clients yet.
As of October 2026.