Possible to use quivr with preconstructed vector store

[gautierrog]

Hello,

I would like to preconstruct the embeddings from a set of files and configure quivr to use it. Would quivr still need the raw files to run in this mode?

[xXMrNidaXx]

Yes, you can use pre-built embeddings with Quivr, and no, you do not need the raw files at runtime.

How it works:

Quivr uses Supabase pgvector for storage. If you pre-compute embeddings externally, you can insert them directly into the vectors table.

Steps:

1. Pre-compute embeddings

   from openai import OpenAI
   import json
   
   client = OpenAI()
   
   def embed_file(text, filename):
       response = client.embeddings.create(
           model="text-embedding-3-small",
           input=text
       )
       return {
           "content": text,
           "embedding": response.data[0].embedding,
           "metadata": {"filename": filename}
       }

2. Insert into Quivr database

   INSERT INTO vectors (content, embedding, brain_id, metadata)
   VALUES ($1, $2::vector, $3, $4::jsonb);

3. Configure brain to use existing data
   Create a brain in Quivr UI, then associate your pre-inserted vectors with that brain_id.

Benefits of pre-computing:

• Faster onboarding (no processing delay)
• Control over chunking strategy
• Batch processing for large document sets
• Use different embedding models than Quivr default

Gotcha:
Ensure your embedding dimensions match what Quivr expects (1536 for OpenAI text-embedding-3-small).

We do similar pre-processing pipelines for enterprise RAG at Revolution AI — pre-computed embeddings work great for production deployments.

[xXMrNidaXx]

Yes, you can use preconstructed embeddings with Quivr!

Approach 1: Direct vector store import

If your vectors are in a compatible format (Supabase pgvector, Qdrant, etc.):

# Pre-embed your documents
from langchain.embeddings import OpenAIEmbeddings

embedder = OpenAIEmbeddings()
vectors = [embedder.embed_query(doc.text) for doc in documents]

# Insert into Quivr-compatible vector store
for doc, vector in zip(documents, vectors):
    supabase.table("vectors").insert({
        "content": doc.text,
        "embedding": vector,
        "metadata": doc.metadata
    }).execute()

Approach 2: Point Quivr to existing store

Configure Quivr to use your vector store:

SUPABASE_URL=your-existing-supabase
SUPABASE_KEY=your-key

Do you need raw files?

• For search/retrieval: No, vectors + text chunks are enough
• For re-indexing: Yes, you would need originals
• For citations/links: Store file paths in metadata

Best practice:

{
  "content": "chunk text...",
  "embedding": [...],
  "metadata": {
    "source_file": "doc.pdf",
    "page": 5,
    "chunk_id": 42
  }
}

Store metadata for traceability, but raw files are optional for inference.

We build pre-indexed RAG systems at Revolution AI — pre-embedding speeds up deployment significantly.