KOL RAG Search

How the search works

A Retrieval-Augmented Generation (RAG) pipeline: fast lexical + semantic retrieval narrows 100 KOLs to a shortlist, then an LLM ranks & explains each match. Below is each step with the fields it uses (inputs → outputs) and a worked example of real data flowing through it.

Example query: “หาคนรีวิวคาเฟ่ในเชียงใหม่” (find café reviewers in Chiang Mai)

Index build (offline, one-time)

Each KOL’s raw profile + post transcripts are flattened into one searchable text and turned into a 512-dim semantic vector. Display data is attached too.

Model: text-embedding-3-small (512-dim, chunk + mean-pool)

namenicknamebiographycategoriestypescontentGroupsbrandssponsorBrandshometowncurrentgenderageGroupplatformsfollowerspost transcripts → searchText→ embedding→ channels/avatar/posts

Example — KOL #239 “นัดรีวิว” searchText: “นัดรีวิว เพจรีวิวอาหาร คาเฟ่ บุฟเฟต์ ร้านใหม่ | categories: food_and_dining, cooking | platforms: facebook, tiktok, instagram …” → embedding = [0.021, −0.044, 0.118, … ] (512 numbers)

Translate & normalize the query

The query is translated (TH↔EN) so both languages match the same content, then normalized — lowercased, Latin accents and Thai tone marks stripped (café→cafe, คาเฟ่→คาเฟ) for typo/variant tolerance.

query→ normalized + translated query

Example “หาคนรีวิวคาเฟ่ในเชียงใหม่” → translate: + “cafe coffee shop reviewer chiang mai” → normalize: คาเฟ่→คาเฟ, café→cafe (tone marks & accents stripped)

Hybrid retrieval → top 25 candidates

Each KOL is scored by combining semantic similarity with lexical overlap, plus boosts for exact attribute hits (category, platform, region) and a follower tiebreak.

score = 0.65·cosine(embedding) + 0.35·lexical(Thai char-trigrams) + attributeBoost + followerTiebreak

query embeddingsearchTextcategoriesplatformsbrandsnamefollowers → 25 candidates

Example — scoring KOL #239 0.65×0.71 (cosine) + 0.35×0.42 (lexical) + 0.12 (food-category boost) = 0.73 → ranks #1 of 100, kept in the top-25 shortlist

Classify intent (in parallel)

An LLM reads the query and returns its intents (brand / sponsor / location / demographic / platform / composite…) plus keywords. Brand names are expanded to sub-brands — e.g. Nestlé → Nescafé, KitKat, Milo, Maggi — so transcript mentions are found.

Model: gpt-4o

query→ intents[]→ keywords[]

Example output intents = ["content", "location"] · keywords = ["คาเฟ่", "กาแฟ", "เชียงใหม่"]
For “Nestlé” it would expand to ["Nestlé","KitKat","Milo","Nescafé","Maggi"]

Rank & explain (the “generation” in RAG)

The ranker LLM receives the candidate profiles plus a keyword-anchored snippet from each, and the detected intents drive a dynamic prompt. It returns the ordered matches, a 0–100 match score, and a plain-language reason. A deterministic brand-hit guarantee force-includes any KOL who actually mentions the queried brand.

Model: openai/gpt-oss-120b-Turbo (DeepInfra) · fallback gpt-4o

idnamenicknamebiographycategoriesbrandssponsorBrandshometown/currentsnippetintentskeywords → matchPercent→ whyMatch

Example — what the ranker sees & returns for #239 in: { id:239, name:"นัดรีวิว", categories:["food","cooking"], snippet:"…คาเฟ่เปิดใหม่แบบสดๆร้อนๆ แถวพร้อมพงษ์…" } → out: matchPercent: 95, whyMatch: "เป็นเพจรีวิวอาหารที่เน้นคาเฟ่และร้านเปิดใหม่"

Join display data & report cost

Ranked matches are joined to avatar, per-platform channels, sample posts and readable category labels for this page. Token usage is priced per call, and the result (with cost) is pushed to Telegram.

matchesavatarchannelspostscategoryLabels → display cards→ cost (USD)→ Telegram

Example — final card for #239 avatar ✔ · channels [Facebook 556K · TikTok 131K · IG 79K] · labels [Food & Dining, Cooking & Chef] · 3 posts tagged MATCH (caption contains “คาเฟ่”) · cost $0.0029 → Telegram ✔

How v2 works — dimensions instead of one embedding

v1 embeds each KOL as one mean-pooled 512-dim vector and retrieves by cosine — a creator’s few on-topic posts get washed out, and it can’t tell doing from selling. v2 replaces that vector with structured per-post dimensions extracted offline, a GPT-4o router that decomposes the query, an exact dimension match, and a content-only LLM rerank. No embeddings at query time.

Offline dimension extraction (one-time, replaces the embedding)

Every post is read from its RAW signals — full transcript + title + description + OCR — and distilled into structured, searchable dimensions. The “understanding” happens once, here, per post (not one averaged vector per KOL).

Output: data/post-dimensions.json · 100 KOLs · 3,695 posts

transcripttitledescriptionocr → topic[]→ activity[] (doing vs ขายของ)→ format→ brands[]→ products[]→ location→ subjects[]

Example — แจ็ก แปปโฮ, post #5 { topic:["ตกปลา","ทะเล"], activity:["ตกปลา"], format:"vlog", subjects:["ตกปลา","วิถีประมง"] } → grounds the KOL to fishing as a doer, not a fish-seller

Router (GPT-4o) — decompose the query

One LLM call segments the (space-less) Thai query and splits it into a content niche and profile filters: intents, keywords, ~10 Thai + ~10 English sense-preserving variants, exclude (anti-sense), and facets (platform / followers / gender / age / location / role / brand).

Model: gpt-4o (cacheable per query)

query→ intents[]→ variants[] (TH+EN)→ exclude[]→ facets{}

Example — “ยูทูบเบอร์สายเล่นเกม” variants = ["เล่นเกม","เกม","gaming"] · facets.platform = YOUTUBE
“เดินป่า” → exclude:["เกม","gaming"] (real outdoor ≠ survival game)

Dimension match (no LLM, ~1ms)

Hard gates (gender / age / platform / followers / role) filter the 100 KOLs; then each surviving post’s dimensions are substring-matched against the variants → grounded posts. exclude terms demote wrong-sense posts (selling, survival-game). Brands match in post dimensions (the profile brand list is only a recall fallback).

score = grounded posts (+ name / location / brand signals) · exclude → 0

variantsexcludefacetspost-dimensions→ ranked candidates + grounded post URLs

Example — “ตกปลา” แจ็ก แปปโฮ activity:["ตกปลา"] → 8 grounded posts · a fish-seller activity:["ขายของ"] → demoted by exclude:["ขายปลา"]

Dedupe + LLM rerank (content intent only)

For a content query, each candidate’s distinct dimensions are deduped (no repeated terms) and sent with the query to the LLM, which ranks by genuine niche relevance and writes a Thai whyMatch. This fixes vertical precision (a heavy-promo account no longer tops a beauty query). Name / brand / pure-filter queries skip this — they’re already exact.

Model: gpt-4o · input = deduped dimensions (≤30 terms/KOL), not raw transcripts

querydeduped dimensions→ matchPercent→ whyMatch

Example — “ผู้หญิงสายบิวตี้” a durian-promo account (top by raw count) → dropped; real beauty creators (pantipa.a, Meijimill) rank first

Join display & badge the matched posts

Same display payload as v1 (avatar, channels, categories). The exact posts v2 matched are badged by URL — so the “MATCH” posts are the ones whose dimensions hit the query — and the real matched fields are shown (Posts / Brand / Platform / …).

matchesgrounded post URLsmatched fields→ display cards (correct posts badged)

v1 vs v2 v1 = one embedding + cosine + LLM rank · v2 = per-post dimensions + router + exact match + content-gated rerank, grounded to specific posts

🔎 KOL RAG Search