I’m working with Genkit’s retriever in Firebase , aiming to filter a collection before applying an embedding match. Specifically, I’d like to filter documents where uid equals a specific value (e.g., 123) before performing the nearest neighbour search on embeddings.
The typical way to filter a Firestore collection is like this:
where('uid', '==', 'my uid')
Here's my current setup for the retriever based on Genkit documentation. I may be missing something, or perhaps pre-filtering isn’t supported? It seems essential to be able to filter down to user-specific documents before executing the similarity search.
const firestoreRetriever = defineFirestoreRetriever({
name: 'firestore-articles',
firestore: getFirestore(),
collection: 'articles',
contentField: 'fullText',
vectorField: 'embeddings',
embedder: textEmbeddingGecko,
distanceMeasure: 'COSINE',
metadataFields: ['path']
});
This retriever is being called as follows:
const docs = await retrieve({
retriever: firestoreRetriever,
query,
options: {
k: 1,
limit: 10,
},
});
Does Genkit provide a method for applying a pre-filter on a specific field (e.g., uid) within the defineFirestoreRetriever function, or would this require a workaround? Any guidance or alternative approaches would be appreciated.
Have you tried:
const docs = await retrieve({
retriever: firestoreRetriever,
query,
options: {
k: 1,
limit: 10,
where: {
uuid: '123',
},
},
});
The where option is mentioned in the available retriever options for Firestore vector store: https://firebase.google.com/docs/genkit/plugins/firebase#retrievers