Skip to content

SKILL.md

Latest commit

 

History

History
98 lines (78 loc) · 3.98 KB

SKILL.md

File metadata and controls

98 lines (78 loc) · 3.98 KB
name search-concepts
summary Couchbase Full-Text Search and Vector Search concepts — index prerequisites, query types, hybrid search, RYOW consistency, Flex Index, RAG pipeline, and pagination
description Couchbase Full-Text Search and Vector Search concepts — index prerequisites, query types, hybrid search, RYOW consistency, Flex Index, RAG pipeline, and pagination
metadata
last_verified min_server_version handoff
2026-05
7.0
condition type skill
user asks for language-specific search code
variant
search-python
condition skill
user asks about Couchbase fundamentals or core concepts
getting-started
condition skill
user asks about SQL++ queries or the SEARCH() function
sqlpp-language
condition skill
user asks about document structure or data modeling for search
server-data-modeling

Language routing: The language-specific handoff in the frontmatter above is a variant edge. If the user's language is known, route directly to the matching search-<lang> skill. If unknown, ask before routing.

Couchbase Search — Concepts

Couchbase provides two complementary search capabilities built on the same index and query infrastructure:

  • Full-Text Search (FTS) — text analysis, fuzzy matching, phrase search, geospatial
  • Vector Search — nearest-neighbour semantic search over embedding vectors

Both use the same index service (port 8094), the same SDK API (cluster.search() / cluster.searchQuery()), and can be combined in a single hybrid search query.

For index prerequisites, query types, hybrid search, RYOW consistency, Flex Index, and pagination, see shared/server/search-concepts.md.

RAG Pipeline

Retrieval-Augmented Generation (RAG) uses Couchbase vector search to retrieve relevant documents and pass them as context to a language model.

The 5-step pattern

1. CHUNK    — split source text into passages (e.g. 512 tokens with overlap)
2. EMBED    — generate a vector for each chunk using an embedding model
3. STORE    — upsert each chunk as a Couchbase document with an "embedding" field
4. RETRIEVE — at query time, embed the user's question and run a vector search
5. AUGMENT  — pass the top-K retrieved chunks as context to the LLM prompt

Store step (language-agnostic pseudocode)

for chunk in chunks:
    vector = embedding_model.encode(chunk.text)   # float[] of length = dims
    doc = {
        "text": chunk.text,
        "source": chunk.source,
        "embedding": vector
    }
    collection.upsert(f"chunk::{chunk.id}", doc)

Retrieve step (language-agnostic pseudocode)

query_vector = embedding_model.encode(user_question)
results = cluster.search(
    index_name,
    VectorSearch(VectorQuery("embedding", query_vector, num_candidates=10)),
    SearchOptions(fields=["text", "source"], limit=5)
)
context = [row.fields["text"] for row in results.rows()]

Key constraints

  • dims in the vector index definition must exactly match the embedding model output size (e.g. 1536 for text-embedding-ada-002, 768 for all-MiniLM-L6-v2).
  • Use templates/vector-index.json as the index definition template — set dims to match your model.
  • Vectors must be stored as a JSON array of floats, not base64.

Framework integrations

Couchbase has official integrations for both major RAG frameworks:

Both integrations handle index creation, document upsert, and similarity search. Use them when building production RAG pipelines rather than calling the SDK directly.