RAG Pipeline

Full pipeline: search for relevant chunks → build prompt → generate answer via LLM.

Usage

from sententia.llm.openai import OpenaiProvider
from sententia.index import Index
from sententia.storage import Storage
from sententia.rag import ask

storage = Storage("/data/markdown")
index = Index(storage, "/data/index.faiss")
provider = OpenaiProvider("https://api.openai.com", "gpt-4", "sk-...")
result = ask("how to set up authentication?", index, provider, top=10)
# {"answer": "To set up authentication...", "sources": ["docs/auth.md", "docs/setup.md"]}

Parameters

Parameter	Type	Default	Description
query	str	—	User question text
index	Index	—	Index instance for chunk search
llm_provider	Provider	—	Configured LLM provider
top	int	10	Number of context chunks

Result

{
    "answer": "Generated answer text",
    "sources": ["docs/auth.md", "docs/setup.md"]
}

• answer — generated answer text
• sources — list of unique source file paths (order preserved)

Algorithm

1. Find relevant chunks via index.search(query, top)
2. Join chunk texts with double newline
3. Build prompt:
4. Opening: "Use the following context to answer the question:"
5. Context block: chunk texts wrapped in triple backticks
6. Question block: query wrapped in triple backticks
7. Instruction: if context is insufficient, respond "Insufficient data to answer this question"
8. Generate answer via llm_provider.generate(prompt)
9. Extract unique source file paths

Insufficient Data

If the context is insufficient to answer, the LLM will return:

"Insufficient data to answer this question"