Talk with Lex

​

Key features

• Passages as documents: Each document is about one minute of speech (110 to 190 words) that never mixes two speakers. It keeps its episode, guest, speaker, chapter, and start time, so every result can point to youtube.com/watch?v=…&t=…s.
• Exact phrases and typos: Wrapping a query in double quotes, such as "love is the answer", runs a phrase search that returns the 7 moments where it was said. A misspelled Dostoyevski still finds Dostoevsky thanks to typo tolerance.
• Words to meaning slider: The slider sets the semanticRatio of a hybrid search. The embedder is a HuggingFace model (BAAI/bge-small-en-v1.5) running inside Meilisearch, so a query like “afraid to die” finds conversations about mortality that never use those words.
• Who said it: A filter on the isLex field keeps only Lex’s own words or only his guests, and facet counts show how many passages each side has. A guest list uses facet search to type-ahead across more than 350 guests.
• One result per episode: A toggle adds distinct: "episodeId" to the query, so each episode appears once with its best passage. Because facets are then counted per episode, a second query without distinct in the same multi-search request provides the “+18 more moments in this episode” counts. See distinct attribute.
• Conversation around a hit: Expanding a result filters on the same episodeId and a position range, sorted by position, to show the passages just before and after.
• Chat with the archive: The chat page uses the chats API. The model searches the transcripts on its own, often several times per question, and cites the second each claim comes from. The interface lists the searches it ran and the passages it read, as described in display source documents.
• Chat with one episode: The same chat, restricted to a single episode by a tenant token whose search rule filters on episodeId. The model cannot retrieve passages from any other episode.

​

Example search and conversation

1. Search meaning of life and choose “Lex”: Returns only the questions Lex asked about the meaning of life, each with the timestamp and the chapter it belongs to.
2. Turn on “One per episode”: The list switches from passages to episodes, and each result shows how many other matching moments its episode contains.
3. Ask the chat “What does Lex think love is?”: The model runs searches such as “what is love” and “love is the answer”, each with the filter isLex = true, then answers with links that jump to the moments it quotes.

​

How it works

import { generateTenantToken } from "meilisearch/token";

// Parent key: actions ["search", "chatCompletions"] on the chunks index only
const token = await generateTenantToken({
  apiKey: process.env.MEILI_CHAT_KEY,
  apiKeyUid: process.env.MEILI_CHAT_KEY_UID,
  searchRules: {
    "lex-chunks": episodeId ? { filter: `episodeId = "${episodeId}"` } : {},
  },
  expiresAt: new Date(Date.now() + 30 * 60 * 1000),
});

// The browser then calls POST /chats/talk-with-lex/chat/completions with this token

​

Links