MIG is the picture Ask Medwiki builds of each person from the questions they ask.
It reads every question and answer, notes what they reveal about the user — who they are,
what condition they are dealing with, what they misunderstand, what they might buy — and keeps
one running profile per conversation and one per user.
It listens, it doesn't interruptThe profile is inferred from normal conversations. One optional question is shown while each answer loads.
Runs in the backgroundOne extra model call per turn, after the answer is sent. It never slows down or breaks a reply.
Every guess has a confidenceA trait is a guess with a 0–100% score until the user states it themselves.
Two profilesOne for each conversation, and one for the user across all their conversations.
How it works
* Feeding the profile into answers is switched off by default and enabled partner by partner (ASK_MEDWIKI_MIG_CONTEXT).
Audience parameters are stored only — nothing in Nucleus delivers ads today.
What MIG keeps about a user
Stable traits
Don't change between questions. Each turn that points the same way adds confidence: 53% after one signal, 71% after two, 89% after three, capped at 95%.
SexAge cohortLiteracy levelLanguageCode mixing
Evolving states
How the user is right now. The latest turn wins. If a state isn't seen again it fades each turn and is dropped after about four turns.
What the user told us in reply to a profile question. Held at 100% and no later inference can change it.
Age groupGenderAsking forDiagnosed conditionWhere they liveDaily medicinesWhere they buyBrand or genericDoctor visitsTreatment preferenceHealth insurance
Running lists and scores
Everything mentioned so far is collected, and an adherence risk score (0 to 1) moves up or down by at most 0.2 per turn, with a trend.
The user as an audience, from fixed lists of values so they work as segments. Interests are counted; the rest keep the latest value. Never sent to the answer model.
Separately, the session records browser and source details the client sends — campaign (UTM) tags, referrer, device, locale, place, analytics IDs and an ad_consent flag. These are facts about the visit, not inferences.
Worked example — one user, two conversations
An illustrative user, not a real one: Sunita, who has had type 2 diabetes for three years and asks on WhatsApp in Hindi written in English letters. The left column is the chat; the right column is what MIG takes from each turn.
This is a best case, written to show every kind of signal. The model is usually more cautious than this: in test runs it often left sex blank despite feminine verb forms, and read a "how much does the generic cost" question as product interest only, with no purchase intent. Expect real profiles to fill more slowly and with more blanks.
A profile question only goes out with a health question — never with a greeting, "how does this work?" or a message that is turned away. The age group is always asked first; after that the question is picked to fit what was just asked (a medicine question brings "Do you take any medicines every day?", a disease question brings the diagnosed-conditions one), so the order below is one possible order, not a fixed one.
Conversation 1
WhatsApp · 7 turns (4 questions, 3 profile answers) · linked to her phone number
Turn 1
Mujhe 3 saal se sugar hai, Glycomet 500 leti hoon. Ab sugar normal aa rahi hai to dawa band kar sakti hoon?
I've had diabetes for 3 years and take Glycomet 500. My sugar is normal now, so can I stop the medicine?
Sent while the answer is preparedWhich age group are you in? 1. Under 18 2. 18–39 3. 40–59 4. 60 or above
MIG picks up
Sex: female 53%Language: HindiDisease: type 2 diabetesMedicine: Glycomet 500Decision stage: managingMisunderstanding: stop meds when normalCondition duration: long termInterest: diabetes careRisk +0.15
"leti hoon" and "sakti hoon" are feminine verb forms, which the model may take as a first signal for sex — it does not always. Wanting to stop because readings are normal is a known misunderstanding.
Turn 2 — profile answer
3
MIG picks up
Stated: age group 40–59Age cohort: adult 100%
Saved as its own turn, with no model call, and kept out of the medical Q&A history. She could have sent it before or after the answer arrived.
Turn 3
Padosan ne bataya karele ka juice peene se sugar jad se khatam ho jaati hai, WhatsApp pe bhi aaya tha. Sahi hai kya?
My neighbour said bitter gourd juice cures diabetes completely, I saw it on WhatsApp too. Is that right?
Sent while the answer is preparedWhat is your gender? 1. Female 2. Male 3. Prefer not to say
A sticky flag: it stays on the profile even if she never mentions WhatsApp forwards again.
Turn 4 — profile answer
1
MIG picks up
Stated: gender FemaleSex: female 100%
The 53% guess is replaced by a stated fact.
Turn 5
Kabhi kabhi subah ki goli bhool jaati hoon. Is dawa se pet bhi kharab rehta hai.
Sometimes I forget the morning tablet. This medicine also upsets my stomach.
Sent while the answer is preparedWho are you mostly asking these questions for? 1. Myself 2. My parent 3. My child 4. My husband or wife 5. Someone else
Asking about price or availability is the cue for "ready to buy", but the model often stops at "researching" or records only the product interest. Nothing new on the health side.
Turn 7 — profile answer
1
MIG picks up
Stated: diagnosed condition Diabetes
Confirms what MIG had already inferred from turn 1.
Conversation 2
WhatsApp · a week later · 4 turns · same phone number, so it joins the same user
Turn 1
HbA1c 8.2 aaya hai. Ye zyada hai kya?
My HbA1c came back as 8.2. Is that high?
Sent while the answer is preparedWhere do you live? 1. A big city (metro) 2. A smaller city or town 3. A village
Latest intent is "researching", but the profile also keeps the peak ("ready to buy") because one turn alone is noisy.
Turn 4 — profile answer
1
MIG picks up
Stated: takes medicines every day
Five of the eleven questions answered and one ignored after two conversations.
Confidence that Sunita is female
Without the question, a second signal would have reached 71% and a third 89%.
Adherence risk score over her six questions
An ideal user profile
S
Sunita — user profile after both conversations2 conversations · 11 turns (6 questions, 5 profile answers) · illustrative
This is what a well-filled profile looks like: the basics are stated rather than guessed, the condition and medicines are named, the misunderstandings are specific, and there is a clear buying signal.
Health profile
Female
told us
Adult
told us
Medium
95%
Hindi
95%
Moderate
95%
Type 2 diabetes
90%
Managing
80%
Calm
60%
Quantified
65% · fading
dropped
faded
Low
60%
Poor
60%
stated by the user inferred, with confidencePercentages follow the accumulation rules; whether a signal is picked up at all is the model's call."Fading" = an evolving state not seen in the last turns. Side-effect tolerance was last seen five turns ago, so it has been dropped.
Blank fields are normal — MIG leaves a field empty rather than guess without a signal. Five questions are still to come in later conversations (where she buys, brand or generic, doctor visits, treatment preference, insurance). "Who are you asking for" was shown and ignored, so it is not asked again.
What the answer model would see
When the context layer is switched on for a partner, this short block is added to the answer prompt. It contains only traits at 50% confidence or more, the medicines, symptoms and misunderstandings, and the stated facts. Audience parameters are never included.
USER PROFILE (MIG) — inferred from this user's earlier Ask Medwiki questions. It may be wrong.
Use it only to pitch the tone, depth, language and examples of your answer. Do not mention the
profile, do not state any of it as a fact about the user, and always go by what the user says
now if it differs.
- sex: female
- age_cohort: adult
- literacy_level: medium
- language: hi
- code_mix_level: moderate
- disease: Type 2 Diabetes
- decision_stage: managing
- emotional_state: calm
- severity_awareness: quantified
- self_med_risk: low
- info_source_quality: poor
- medicines: Glycomet 500
- symptoms: stomach upset
- misunderstanding_tags: stop_meds_when_normal, missed_dose_normalization
- age_group (the user told us): 40–59
- diagnosed_condition (the user told us): Diabetes
- gender (the user told us): Female
- location_type (the user told us): A smaller city or town
- regular_medicines (the user told us): Yes
What makes a conversation useful to MIG
Rich — the profile fills quickly
Says who the question is for ("meri mummy ko…", "I have…").
Names the condition and how long it has been there.
Names the medicine and dose actually being taken.
Describes what they really do — skipped doses, stopping, changing the dose themselves.
Quotes numbers from reports (HbA1c 8.2, BP 150/95).
Mentions where a belief came from — doctor, family, WhatsApp.
Asks about price, substitutes or where to buy.
Runs to several turns and answers the profile questions.
No sex, age, condition, risk or buying signal. That is fine — MIG leaves those blank. One-word lookups and single-turn chats produce profiles like this, which is the reason the profile questions exist.
How a profile is kept honest
Kind of field
How new turns change it
Example
Stable trait
The value seen most often wins; confidence grows with each agreeing turn, up to 95%.
Sex 53% → 71% → 89%
Evolving state
The latest turn replaces it. Unseen, it fades each turn and is dropped after about four.
Confused in turn 3, calm by the next conversation
Sticky risk flag
Can only stay or get more serious. Confidence grows with every signal.
Info source "poor" stays after one WhatsApp cure
Stated fact
Fixed at 100%. A later guess cannot overrule it.
Age group 40–59 pins age cohort to adult
Lists
Only ever added to.
Medicines, symptoms, misunderstandings
Interests
Counted, so the most frequent topic rises to the top.
Diabetes care ×6
Purchase intent
Latest value, plus the strongest ever seen.
Researching · peak: ready to buy
Limits and safeguards
It can be wrong. Everything except a stated fact is a model's inference. Read the confidence, not just the value.
Identity is unverified. Conversations are grouped by phone number, then device or partner user ID. A shared phone or device mixes two people into one profile.
Answers are unaffected by default. The profile reaches the answer model only for partners listed in ASK_MEDWIKI_MIG_CONTEXT.
Minors. A user who says they are under 18 is asked no further profile questions and is marked as excluded from audiences.
Advertising. Audience parameters and ad_consent are only recorded. Nothing delivers ads or checks consent yet; any future use must honour consent and the rules for health data.
Question wording. A profile question is asked only in the language of the answer, and only once that language's wording — the question and the three messages around it — has been approved on the MIG Questions page. A language with unapproved messages gets no questions, rather than a question with an English intro.
Who can see it. Profiles are visible only to logged-in Nucleus users.
Where to find it in Nucleus
Conversations — open any conversation to see its MIG card. Switch between "This conversation" and "User (all conversations)", and use Generate for older conversations.
MIG Questions — edit and approve the wording of the profile questions in each language.
Partners — turn profile questions on or off for a partner.