Chat Everywhere

A full-screen chat with a conversation list, a side panel from the top bar on every screen, streaming answers, live progress, example questions and 👍/👎 feedback.

Answers With Sources

Every answer cites the records it used. One click on a source button opens the real record, so users can check any figure themselves.

Structured Questions

Counts, totals, top-N by customer or salesperson and filtered lists. Query plan for local models, Tool calling for cloud models, or Retrieval only.

Knowledge Sources

Choose the models, a domain and the fields to index. Sources for Contacts, Products, Sales, Purchase, Invoices, CRM, Project and Inventory are created on install when those apps exist; Employees is created switched off.

Hybrid Retrieval

Vector similarity plus PostgreSQL full-text search. pgvector with an HNSW index when available, an optional numpy fallback for small databases.

Attachments

Optionally index the text of files attached to indexed records, using Odoo's own attachment text extraction (attachment_indexation for PDF and Office files).

Always Up to Date

Changed records are picked up every minute, nightly or manually per source. An index monitor shows coverage, stale chunks and the last error.

Restricted Fields

Mark fields that must never leave your server; they are redacted for cloud providers. Optional scrubbing of e-mails, phone and bank numbers.

Usage, Cost & Audit

Every provider call and every refusal is logged. A usage report shows calls, tokens, latency and estimated cost; a daily token cap per user keeps spend under control.

Create & Update From Chat

“Change the phone of Gemini Furniture to …” or “Create a sale order for Azure Interior with 3 Desk Organizer” becomes a change card — fields and document lines, with current and new values and Confirm / Discard buttons. Works with tool calling and with the Ollama query plan mode.

AI Agent Roles

Per group or user: which models, create and/or update, which fields, and which records may be updated (for example only quotations). A role never grants more than the user's own access rights.

Record Changes Report

Every prepared change with user, role, fields and outcome: saved, discarded, failed or expired. Validated again at confirmation, locked against double clicks, never editable.

Give people access

Settings → Users & Companies → Users. The main Administrator receives the AI Agent Administrator role automatically.

Open a user, go to Access Rights and set AI Agent to User or Administrator.

For someone who only reviews what was sent to AI providers, set AI Agent Audit to Auditor.

Save, and ask the user to reload the page so the menu and the top-bar icon appear.

Add a provider — the form adapts to your choice

AI Agent → Configuration → Providers → New. Pick Ollama, OpenAI or Anthropic and only the fields that provider needs are shown.

Ollama: Base URL and Timeout are filled in. On Save the installed models are loaded from your server; if only one model of a kind exists, it is selected for you.

Ollama defaults are tuned for a CPU: Keep-alive 30m, Timeout 300 s and Structured Questions: Query plan. Opening the chat warms the model up in the background.

OpenAI / Anthropic: paste the API key — it is never shown again — or reference it as env:OPENAI_API_KEY or config:openai_api_key to keep it out of the database.

Anthropic offers no embeddings, so pair it with an Ollama or OpenAI Embedding Provider. Tick the default boxes and click Test Connection.

A new Ollama provider

Installed models loaded, Query plan selected

OpenAI

Anthropic, with a separate embedding provider

Check the settings

AI Agent → Configuration → Settings. The defaults are a sensible start.

Vector Backend tells you whether pgvector is in use. Without it, tick numpy Fallback for small databases.

Retrieval: Top-K Chunks, Similarity Threshold and the vector / keyword weights.

Safety: Grounding Verifier, PII Scrubbing, Store Full Prompts — all off until you switch them on.

Limits: Daily Token Cap per user and retention days for conversations and the audit log.

Choose what the agent can read

AI Agent → Configuration → Knowledge Sources. A source is a model plus the fields the agent may read. Adding a source gives nobody extra rights.

Pick the Model and the Trigger mode: every minute, nightly or manual.

Add the Indexed Fields and an optional Domain. For counts and rankings, key fields such as status, amounts, date, customer and salesperson are used automatically.

Add Restricted Fields that must never go to a cloud AI. Fields protected by user groups (for example salaries) are refused.

Keep Allow Structured Queries ticked, and tick Index attachments if files should be read too.

Build and watch the index

AI Agent → Configuration → Index Monitor. Indexing runs in the background; nobody has to wait on a screen.

Follow Eligible and Indexed Records, Chunks, Stale and Coverage (%). A line turns red with the Last Error.

Index Now reprocesses only changed records. Re-index / Purge… offers a full re-index, a purge and an ANN index rebuild.

Changing the embedding model is guarded: a dedicated wizard explains the impact and rebuilds everything, so incompatible vectors are never mixed.

The index monitor

Re-index, purge or rebuild

Let people change records (optional)

AI Agent → Configuration → Roles. Skip this step to keep the agent read-only.

Name the role and choose the Groups (for example Sales / User) or single Users who get it.

Add a permission line per model: Create, Update, the Allowed Fields and, for a document's lines, the Allowed Line Fields (for example Product and Quantity). Read-only, computed and related fields cannot be chosen; technical and security models are refused.

Limit updates with Updatable Records, for example only quotations. The user's record rules still apply on top.

In Settings, Record changes sets how long a prepared change waits for Confirm (60 minutes by default).

A role: groups, models, operations and allowed fields

Reporting → Record Changes

Ask a question

AI Agent → Chat. Type a question and press Enter. A status line shows what is happening: thinking & working, asking the model, querying Odoo data, composing the answer. The chat always scrolls to the newest message.

“What is the email address of Gemini Furniture?” — a lookup with a citation.

“How many quotations do we have?”, “Show the highest sales order”, “Who is our top customer last month?” — computed by Odoo in the user's time zone.

Read the How I searched line: it shows the filters and sorting used, so a misunderstood question is obvious at a glance.

Conversations are saved; follow-up questions keep the context. Start a new conversation when the topic changes.

Ask from any screen, or about one record

No need to leave the work you are doing.

Click the robot icon in the top bar: the chat slides in as a side panel on every screen.

On a record of an indexed model, open ⚙ Actions → Ask AI about this record. The chat opens focused on that record, for example “Summarise this customer”.

Questions about the outside world are refused, and record types that are not indexed get a clear “I don't have access to…” message.

The side panel, declining an off-topic question

Ask AI about this record

Create or update a record from the chat

For users with an AI Agent Role. Start the message with an action word such as Create, Change or Set.

“Change the phone of Gemini Furniture to +32 470 12 34 56” — a change card shows the field, its current value and the new value.

Confirm saves it with the user's own rights; Discard drops it. The change is validated again at that moment.

The answer text is fixed, so the model can never claim a change was saved. Values not written in the message are left out.

“Create a sale order for Azure Interior with 3 Desk Organizer and 5 Desk Pad” — the card lists one row per order line; prices and taxes are computed by Odoo on confirmation. Not possible from the chat: deleting records, confirming or cancelling documents, and changing lines that already exist.

The change card, waiting for confirmation

Saved after Confirm, with a link to the record

Keep an eye on usage, cost and data flow

AI Agent → Reporting. Visible to administrators and auditors.

Usage & Cost: calls, tokens in and out, average latency and estimated cost from the prices you enter, by provider, user or day.

Audit Log: every provider call and refusal, append-only. The Sent to External Provider filter shows exactly what left your server — with Ollama only, it stays empty.

Answer Feedback: every 👍/👎 with the user's comment, to find the questions the setup handles badly.

Record Changes: every change prepared by the agent, with user, role and outcome.

Usage & Cost

The append-only audit log

Local or Cloud — Your Choice

Ollama keeps every byte on your own server, and is always allowed to see restricted fields. Chat and embedding models are read from the server's installed list.

OpenAI and Anthropic answer in seconds. Restricted fields are redacted before sending, and each call is visible in the audit log.

Use different providers for chat and for embeddings, and switch the default at any time.

1. Scope Guardrail

A local rule, no AI call: questions that match none of your indexed models, fields or records are refused. The daily token cap is checked too.

2. Hybrid Retrieval

Vector and full-text candidates from the chunks of the active knowledge sources, limited to the user's companies.

3. ACL Filter

Candidate record IDs are re-checked with an ORM search as the asking user. Records they cannot open are dropped.

4. Redaction

Restricted fields are replaced by [REDACTED] for cloud providers; optional PII scrubbing removes e-mails, phones and bank numbers. Context delimiters inside record text are neutralised.

5. Safe Queries

Query plans and tool calls go through a strict domain parser: known fields only, no group-protected or blocked fields, no SQL, no eval.

6. Answer Checks

Numbers not found in the evidence are withheld; an optional grounding verifier double-checks the answer. Everything is written to the audit log.

Your Policy, Not Ours

Which models and fields are readable, which records and fields each role may change, restricted fields, top-K, similarity threshold, chunk size, conversation history, tool limits, PII scrubbing, prompt storage, daily token cap and retention — all settings, not code. Prompt templates are editable records.

Secure by Design

No sudo() in the answering pipeline — a test enforces it. API keys are write-only. Conversations are private to their owner by record rule. Record changes are saved as the confirming user, never with elevated rights. The test suite includes access-isolation, no-egress and record-change tests.

Fast on Ordinary Hardware

Compact indexed text, keep-alive, model warm-up when the chat opens, and one short AI call per structured question. Measured on a laptop CPU without a GPU (Ollama, llama3.2:3b): answers in about 6–21 seconds.

Answers in the Question's Language

The agent is instructed to reply in the language of the question; quality depends on the model. Interface strings are translatable (a .pot template is included).