How isanna runs.

The shared path

client
   │
   ▼
Cloudflare      DNS · CDN / proxy
   │
   ▼
Fly.io          per-product application containers
   │
   ▼
Postgres        schema-isolated per product

One path to learn rather than one per product: a service grows without being rebuilt, and a new product inherits the route instead of inventing it. Builder is the exception — it is a command-line workflow with nothing hosted at all.

Data and isolation

Each product owns its own schema and connects with its own login role that cannot bypass row-level security. Where a table belongs to a user, the ownership check is a policy in the database rather than a condition someone has to remember to write.

Not every table is owned by a user, and those are scoped by the application instead — a choice each migration records in writing, so which mechanism protects which table is answerable rather than assumed.

Retrieval

Memory for coding agents is a retrieval problem before it is a storage problem. Vector search alone, over a store that keeps growing, returns what is similar rather than what is relevant — and cannot tell a current decision from the one that superseded it. The design follows from that, and it is built in stages.

Running

Hybrid search — lexical and vector together — with mxbai-embed-large producing the embeddings, then a reranker over the merged candidates. First-stage retrieval is allowed to be generous, and the reranker decides what actually surfaces.

Running

The embedding model and its dimension are decoupled from the storage schema, so the model can be replaced without a migration.

Next

Query rewriting, and contextual retrieval — embedding a chunk together with the context that makes it readable on its own.

Later

Hierarchical summary trees, graph retrieval and late-interaction models such as ColBERT. These are strong where a corpus has heavy structure or needs reasoning across documents; this one is neither yet, so they sit behind the work above rather than in front of it.

The sequence is the decision worth explaining: each stage has to earn the index it adds and the latency it spends, so the cheap wins land before the elaborate ones.