LIA Blog
Discover how LIA works: technical articles, practical guides and insights about your open source personal AI assistant.
Architecture
How LIA orchestrates its AI agents, plans tasks, and keeps humans at the center of decisions.

Architecture4 min read
Pipeline or ReAct: two ways for LIA to handle your request
LIA can plan your whole request ahead of time, or move step by step, reading each result. Here is what each mode does, why I kept both, and how to pick the one that suits you.
reactpipelineexecution-mode

Architecture4 min read
A conductor and specialists: how LIA handles your request
Behind each message, LIA does more than call a model. Here's how it sorts your request, prepares a plan, brings in the right tools in parallel and answers you, and why I chose to split the work this way.
langgraphmulti-agentorchestration

Architecture4 min read
A plan before acting: how LIA prepares and checks your multi-step requests
When your request takes several actions, LIA writes a plan, checks it, then runs it in parallel when it can. Here's what that plan looks like, what's checked before the first call, and what happens when a step fails.
plannerexecution-plantask-orchestration

Architecture4 min read
Staying in control without being bothered: how LIA asks for your approval
An assistant that acts on your behalf has to know when to stop and ask you. Here's how I wanted LIA to do it: at the right moment, on what is really going to happen, and without drowning you in questions.
hitlapprovalhuman-validation

Architecture4 min read
A model for each task, from the provider of your choice
LIA isn't tied to any AI provider. Here's why I wanted you to be able to choose a model per task, how LIA adapts to each provider's habits, and what stays online even with models running on your own machine.
openaianthropicgoogle
Integrations
Connect LIA to Google, Apple, Microsoft, Philips Hue and extend it with MCP and custom skills.

Integrations3 min read
MCP servers you use in several steps: iterative mode
Some MCP servers can't be driven with a single call: you have to read their instructions, try, look at the result, try again. Here's how LIA hands them to a dedicated assistant that works step by step, and when to turn this mode on.
mcpreact-agentiterative

Integrations4 min read
Gmail, Calendar, Drive: what LIA does with your Google account
Connecting your Google account lets LIA read your emails and handle your calendar, contacts, files and tasks in everyday language. Here's what it does with them, what it asks you before acting, and where the limits are.
googlegmailcalendar

Integrations4 min read
iCloud or Microsoft 365: LIA adapts to your tools
Not everyone lives at Google. Here's how LIA connects to iCloud and Microsoft 365, why you can mix providers, and why you phrase your request the same way everywhere.
appleicloudmicrosoft

Integrations4 min read
Plugging your own tools into LIA with MCP
MCP lets you add tools LIA doesn't have, like your ticket tracker or a database. Here's how I opened that door without making you lose control, and what this protocol doesn't guarantee.
mcpmodel-context-protocolextensibility

Integrations4 min read
Skills: teaching LIA your way of doing things
A skill is an instruction sheet that teaches LIA a method, a format or a small program. Here is what it can contain, how to create one just by chatting, and how LIA protects herself from skills written by others.
skillsskill-generatorskill-library
Features
Voice, documents, personal journals, proactive notifications and much more.

Features4 min read
LIA's psyche: a mood that colors the way she talks to you
LIA has a mood, passing emotions and a relationship with you that evolves. Here is what it's for, how it changes the tone of her answers without touching the facts, and how you adjust it.
psycheemotionspersonality

Features3 min read
Watching LIA work: the steps that show up while you wait
A complex request can take a few seconds. Meanwhile, LIA shows you what she's really doing, step by step, and keeps that account under each answer. Here is what you see, and what it doesn't show.
execution-stepsreal-timeprogressive-display

Features4 min read
Talking to LIA when typing is a pain
Dictate a question, listen to the answer, talk in real time or let LIA call you: here are the ways to speak with LIA, what stays on your server and what goes to a provider.
voicespeech-to-texttts

Features4 min read
Your documents, one question away: knowledge spaces
Drop your PDFs, spreadsheets or presentations into a space, and LIA uses them to answer you, citing its sources. Here's how it picks the right passage, where documents can come from, and what it can't read.
ragdocumentshybrid-search

Features4 min read
LIA's journal: what she learns about you, and how you stay in control
LIA keeps a journal where she notes what works with you, and draws a portrait from it. You'll see how she writes it, how it adjusts to your reactions, and how to read it, correct it or switch it off.
journalsintrospectionpersonality

Features4 min read
When LIA takes the initiative: useful notifications that don't overwhelm you
LIA can let you know on her own: rain before your appointment, time to leave, an e-mail waiting. Here is how she decides to speak to you, and above all when she stays quiet.
heartbeatnotificationsproactive

Features4 min read
Reminders and routines: let LIA take care of the timing
A reminder alerts you at the right moment; a routine puts LIA to work at a set time or when something happens. Here is how both work, how they follow your local time, and what they don't do.
remindersscheduled-actionsautomation

Features3 min read
LIA's ephemeral experts: delegating an analysis without losing control
For a question that deserves a specialist's eye, LIA can hand the research to a temporary expert, or to several in parallel. Here is what that expert does, what it isn't allowed to do, and what it costs.
sub-agentsdelegationparallel
Security & Privacy
How LIA protects your data with encryption, human validation and privacy controls informed by GDPR.

Security & Privacy4 min read
Deactivate, delete, erase: what happens to your account when you leave
A LIA account goes through four states, from active to erased. You'll see what each one blocks, what is really deleted, and why a few billing records remain.
accountlifecyclegdpr

Security & Privacy4 min read
When the model makes things up: how LIA guards against it
A language model fills gaps with something plausible: a parameter, a time, an explanation. Here are the safeguards I placed before the action, in the answer and when things fail, so that LIA only presents as true what it received or computed.
hallucinationdefenseparameter-stripping

Security & Privacy4 min read
Protecting your account and your data: LIA's security, layer by layer
LIA reads your e-mails, your calendar and your documents, so she has to be solid. Here are the barriers I put around your account, around what LIA reads and around the server, and what they don't cover.
securitydefense-in-depthrate-limiting

Security & Privacy4 min read
Your data stays yours: what LIA keeps, shows and erases
Where your data lives, who can see it, how to take it with you or have it erased: here is what I wanted for your privacy in LIA, with its limits stated plainly.
privacygdprdata-protection
Technical
Observability, prompt engineering, token optimization and multilingual architecture.

Technical3 min read
Editing the whole model catalogue in a spreadsheet, safely
When a provider revises its entire price list, fixing models one by one becomes a chore. Here is how I made the round trip with Excel possible, and the rules that make a file safe to reimport.
excelimport-exportadmin

Technical4 min read
Finding what matters without dragging in everything: the thresholds of search by meaning
LIA finds your memories, your journal and your documents by their meaning, even when your words differ from theirs. It all comes down to a threshold: too low, it brings back noise; too high, it forgets. Here's how I chose these thresholds.
embeddingssimilaritycalibration

Technical4 min read
Tuning LIA's models without a restart, and without impossible configurations
Changing a task's model should be simple for the administrator and free of surprises for everyone. Here's where these settings live, the safeguards that refuse what the provider would refuse, and the two ceilings that cap spending.
llm-configadmin-uigovernance

Technical4 min read
Seeing what's going on: how LIA watches over itself
When something slows down or breaks, someone has to see it before you do and understand why. Here's how LIA measures, traces and warns, and how it reads its own health.
prometheusgrafanalangfuse

Technical4 min read
LIA's instructions, written and reviewed like code
What the model reads decides what it does. Here is why I put each of LIA's instructions in a versioned file, and the rule that keeps them honest: a prompt only states what the code enforces.
promptsversioningprompt-engineering

Technical3 min read
Showing the model what matters: the services that lighten LIA's work
Every call to a language model costs time and money. Here is how LIA understands your question from the start, shows the planner only the useful tools, and skips the steps whose path is already known.
optimizationcachingpattern-learning

Technical4 min read
Six languages, and yours everywhere in LIA
Translating the menus isn't enough: a confirmation, an email or a notification in the wrong language is all it takes to dent trust. Here's how LIA decides which language to speak to you, and where it still falls short.
i18nmultilinguallocalization
LIA is ready. And slightly impatient.
An assistant that acts, knows you and anticipates, while you stay in control of it.
Create my account