Using the SQLite Store

SQLiteVectorStore persists documents and vectors in a local SQLite database file. It uses exact cosine similarity, which is appropriate for local small-to-medium indexes.

Create an index

from pyaistack import RAG
from pyaistack.chunking import TextChunker
from pyaistack.loaders import DirectoryLoader
from pyaistack.vectorstores import SQLiteVectorStore

documents = DirectoryLoader("/path/to/knowledge/", file_type="text").load()

rag = RAG(
    vector_store=SQLiteVectorStore("knowledge.db"),
    chunker=TextChunker(),
)
rag.add_documents(documents)

Reuse an index `ask()`

Create a new RAG instance with the same database path and embedding model. Do not add the documents again.

from pyaistack import RAG
from pyaistack.vectorstores import SQLiteVectorStore

rag = RAG(
    embedding_model="embeddinggemma",
    llm_model="gemma3:4b",
    vector_store=SQLiteVectorStore("knowledge.db"),
)

answer = rag.ask("What information is available?")
print(answer.text)

Retrieve chunks with `search()`

Use search() when you need the matching chunks, their metadata, and similarity scores without calling the LLM. This is useful for inspecting retrieval results or building a custom response flow.

results = rag.search("What information is available?", top_k=3)

for result in results:
    print(result.score)
    print(result.document.metadata)
    print(result.document.text)

`ask()` vs `search()`

Use ask() when you want PyAIStack to retrieve the chunks and generate a grounded natural-language answer. Use search() when you only need retrieval.

Aspect	rag.search()	rag.ask()
Purpose	Retrieves relevant chunks only	Retrieves chunks and generates an LLM answer
Calls embedding model	Yes	Yes
Calls chat/LLM model	No	Yes
Return value	`tuple[SearchResult, ...]`	`RAGAnswer`
Includes score	Yes, per result	Available in `answer.sources`
Includes metadata	Yes, per result	Available in `answer.sources`
Has `.text` answer	No	Yes, via `answer.text`
Best for	Debugging retrieval, custom UI, source inspection	Normal user-facing question answering
Example	results = rag.search("budget planning")	answer = rag.ask("How should I plan a budget?")

Filter by metadata

Pass metadata_filter to either search() or ask() to limit retrieval to documents with matching metadata. Every supplied key/value must match. A scalar filter matches a scalar value exactly or one value in stored list metadata; a list filter requires every requested value to be present. Filtering happens before cosine scoring.

finance_results = rag.search(
    "How should a monthly budget be reviewed?",
    metadata_filter={"category": "finance"},
)

answer = rag.ask(
    "How should a monthly budget be reviewed?",
    metadata_filter={"category": "finance", "audience": "general"},
)
print(answer.text)

Add metadata while indexing

Metadata filters can only match fields stored during indexing. Add them with LoadedDocument(metadata={"category": "finance"}) or rag.add(..., metadatas=[...]).

No metadata indexed

if no Metadata available or indexed then result will not found even if it is present.

The store binds itself to the first embedding model and vector dimension. To use a different embedding model, clear and rebuild the index:

rag.clear()

For hundreds of thousands of chunks or more, use a dedicated indexed vector database in a future PyAIStack adapter.