Retrieval-Augmented Generation (RAG)

Build a knowledge base that answers questions from your own documents. This tutorial uses the wippy/embeddings module for vector search and the LLM framework for generation.

What You'll Build

A minimal RAG pipeline:

  1. Ingest markdown documents — split into chunks, embed, persist.
  2. Retrieve — vector search returns the most relevant chunks for a query.
  3. Generate — an LLM call uses the retrieved chunks as grounding context.

Prerequisites

  • A database: db.sql.sqlite (includes vec0 support) or db.sql.postgres with the pgvector extension.
  • An LLM provider configured with an embedding model (e.g. text-embedding-3-small) — see LLM Framework.
  • Wippy project bootstrapped (wippy init, wippy add wippy/embeddings).

Dependencies

Declare the wippy/embeddings dependency and point it at your database. The target_db parameter is the registry ID of the database entry the embeddings table will live in:

version: "1.0"
namespace: app

entries:
  - name: db
    kind: db.sql.sqlite
    file: ./data/app.db
    lifecycle:
      auto_start: true

  - name: embeddings
    kind: ns.dependency
    component: wippy/embeddings
    version: "*"
    parameters:
      - name: target_db
        value: app:db

wippy/embeddings pulls in wippy/llm and the migration that creates the embeddings_512 table (PostgreSQL pgvector or SQLite vec0 virtual table).

Ingest Documents

Splitting is handled by the text module; embedding and persistence by the embeddings library.

-- app/ingest.lua
local text = require("text")
local embeddings = require("embeddings")
local uuid = require("uuid")

local function ingest(doc_id, title, markdown)
    local splitter, err = text.splitter.markdown({
        chunk_size = 800,
        chunk_overlap = 100,
        heading_hierarchy = true,
        code_blocks = true,
    })
    if err then return nil, err end

    local chunks, split_err = splitter:split_text(markdown)
    if split_err then return nil, split_err end

    local batch = {}
    for i, chunk in ipairs(chunks) do
        table.insert(batch, {
            content = chunk,
            content_type = "doc_chunk",
            origin_id = doc_id,
            context_id = tostring(i),
            meta = { title = title, chunk = i },
        })
    end

    return embeddings.add_batch(batch)
end

return { ingest = ingest }

Register the function and its imports:

- name: ingest
  kind: function.lua
  source: file://app/ingest.lua
  method: ingest
  modules:
    - text
    - uuid
  imports:
    embeddings: wippy.embeddings:embeddings

Key points:

  • origin_id groups chunks that belong to the same source document.
  • context_id is an optional sub-key (section, page, chunk index).
  • add_batch auto-splits if total tokens exceed the 8000-token request limit.

Retrieve

Vector search returns the most similar chunks to the query, along with similarity scores:

local embeddings = require("embeddings")

local results, err = embeddings.search("how do I configure TLS?", {
    content_type = "doc_chunk",
    limit = 5,
})

-- results[i].content, .similarity, .meta, .origin_id, .context_id

Filter by origin when you want to ground the answer in a specific document:

local hits = embeddings.find_by_origin("refund policy", "doc-42", { limit = 3 })

Generate an Answer

Compose the retrieved chunks into a prompt and call the LLM. Here the retrieved text is appended to the system prompt; the user's question becomes the user turn:

-- app/answer.lua
local embeddings = require("embeddings")
local llm = require("llm")
local prompt = require("prompt")

local SYSTEM = [[
Answer using only the provided context. If the context does not contain
the answer, say you don't know. Cite the chunk title for each claim.
]]

local function format_context(hits)
    local parts = {}
    for i, h in ipairs(hits) do
        local title = h.meta and h.meta.title or h.origin_id
        table.insert(parts,
            string.format("[%d] %s\n%s", i, title, h.content))
    end
    return table.concat(parts, "\n\n")
end

local function answer(question)
    local hits, err = embeddings.search(question, { limit = 4 })
    if err then return nil, err end

    local p = prompt.new()
    p:add_system(SYSTEM)
    p:add_system("Context:\n\n" .. format_context(hits))
    p:add_user(question)

    local response, gen_err = llm.generate(p, { model = "gpt-4o-mini" })
    if gen_err then return nil, gen_err end

    return {
        answer = response.result,
        sources = hits,
    }
end

return { answer = answer }

- name: answer
  kind: function.lua
  source: file://app/answer.lua
  method: answer
  imports:
    embeddings: wippy.embeddings:embeddings
    llm: wippy.llm:llm
    prompt: wippy.llm:prompt

End-to-End Example

Putting it together behind an HTTP endpoint:

version: "1.0"
namespace: app

entries:
  - name: db
    kind: db.sql.sqlite
    file: ./data/app.db
    lifecycle:
      auto_start: true

  - name: embeddings
    kind: ns.dependency
    component: wippy/embeddings
    version: "*"
    parameters:
      - name: target_db
        value: app:db

  - name: ingest
    kind: function.lua
    source: file://app/ingest.lua
    method: ingest
    modules:
      - text
      - uuid
    imports:
      embeddings: wippy.embeddings:embeddings

  - name: answer
    kind: function.lua
    source: file://app/answer.lua
    method: answer
    imports:
      embeddings: wippy.embeddings:embeddings
      llm: wippy.llm:llm
      prompt: wippy.llm:prompt

  - name: gateway
    kind: http.service
    addr: ":8080"
    lifecycle:
      auto_start: true

  - name: api
    kind: http.router
    meta:
      server: app:gateway
    prefix: /api

  - name: ask
    kind: http.endpoint
    meta:
      router: app:api
    method: POST
    path: /ask
    func: app:answer_http

  - name: answer_http
    kind: function.lua
    source: file://app/answer_http.lua
    method: handler
    modules:
      - http
    imports:
      answer: app:answer

-- app/answer_http.lua
local http = require("http")
local answer = require("answer")

local function handler()
    local req = http.request()
    local res = http.response()

    local body, err = req:body_json()
    if err or not body or not body.question then
        res:set_status(http.STATUS.BAD_REQUEST)
        res:write_json({ error = "question is required" })
        return
    end

    local result, ans_err = answer.answer(body.question)
    if ans_err then
        res:set_status(http.STATUS.INTERNAL_ERROR)
        res:write_json({ error = ans_err })
        return
    end

    res:write_json(result)
end

return { handler = handler }

Seed the index by calling ingest from a setup process or a CLI command (process.lua with meta.command), then query:

curl -X POST http://localhost:8080/api/ask \
    -H 'Content-Type: application/json' \
    -d '{"question":"how do I configure TLS?"}'

Operational Notes

  • Chunk size: 500–1000 tokens is a good starting point. Too small loses local context; too large dilutes similarity scores. Use chunk_overlap (~10–20% of chunk size) to preserve sentences across boundaries.
  • Content types: Use distinct content_type values (doc_chunk, faq, code_snippet) so search can filter by type.
  • Re-indexing: Delete and re-ingest per document via embedding_repo.delete_by_origin(doc_id) before adding new chunks.
  • Hybrid search: For exact-term recall (names, IDs), combine vector search with full-text search over your source table and re-rank.
  • Model choice: The default 512-dimension text-embedding-3-small is cost-efficient. Upgrade to 1024 or 3072 dimensions only if recall is insufficient — bigger vectors mean bigger storage and slower search.

Next Steps