Use cases

SrvDB fits anywhere an application needs similarity search over embeddings, inside its own process, on local disk. Each use case below says why SrvDB fits and which index mode to start with.

1. Local RAG over private documents

Answer questions from PDFs, notes, wikis or tickets with a local LLM, without sending documents anywhere. Chunks are embedded, stored with their text as metadata, and the closest chunks become the LLM’s context.

• Why SrvDB: no server to run, data stays on the machine, upsert makes re-indexing changed files simple, persist() makes the index crash-safe.
• Start with: flat; hnsw for very large corpora.
• Guide: Build a local RAG app.

2. Long-term memory for AI agents

Give an assistant memory across sessions: store facts, past messages and tool results as embeddings, recall the relevant ones each turn, forget with delete.

memory = srvdb.Database("./agent-memory", dimension=384)

def remember(key, text):
    memory.upsert([key], [embed(text)], [json.dumps({"text": text})])
    memory.persist()

def recall(query, k=5):
    return [json.loads(memory.get(i))["text"] for i, _ in memory.search(embed(query), k)]

• Why SrvDB: in-process (no latency to a service), safe across crashes, one directory per user or per agent.
• Start with: flat.

3. Semantic search in desktop and CLI apps

“Search by meaning” in a notes app, email client, file manager or terminal tool, shipped as part of the app.

• Why SrvDB: pip install srvdb with no runtime dependencies, wheels for Linux, macOS and Windows; nothing for users to install or configure.
• Start with: flat, or auto to switch to HNSW as the data grows.

4. Offline and air-gapped systems

Healthcare, legal, finance, defence and on-premises deployments where data must not leave the machine or network.

• Why SrvDB: never opens a network connection, stores everything in one local directory, logs no vector or metadata contents.
• Start with: whichever mode fits the data size.

5. Edge devices and small machines

Search on ARM boards, laptops and embedded Linux where memory and disk are limited.

• Why SrvDB: prebuilt Linux aarch64 wheels; SQ8 stores vectors 4× smaller, PQ up to 32× smaller.
• Start with: sq8; measure recall against flat before trying pq.

6. Code search

Find functions by what they do, not their names, across a local codebase, using a code embedding model.

• Why SrvDB: index a repository locally and keep it current with upsert and delete as files change.
• Start with: flat or hnsw.

7. Image, audio and multimodal similarity

Search photos by text (“sunset over water”) or find similar images, with CLIP-style embeddings; the same works for audio embeddings.

• Why SrvDB: any embedding with 128 to 4096 dimensions works, so the text and image vectors of a multimodal model share one database.
• Start with: flat; hnsw for large photo libraries.

8. Deduplication and near-duplicate detection

Find documents, support tickets or records that say the same thing in different words.

for doc_id, vector in items:
    hits = db.search(vector, k=2)          # the best hit is the item itself
    if len(hits) > 1 and hits[1][1] > 0.95:
        print(doc_id, "looks like", hits[1][0])

• Why SrvDB: scores are true cosine similarities, so a threshold means the same thing in every mode.
• Start with: flat for exact scores.

9. Recommendations (“more like this”)

Recommend articles, products or songs similar to what a user just viewed, from a local catalogue.

• Why SrvDB: a single search call on the item’s own vector; search_batch for many items at once.
• Start with: hnsw for large catalogues.

10. Testing and prototyping RAG pipelines

Run retrieval in unit tests, notebooks and CI without starting a database service.

• Why SrvDB: a database is just a temporary directory; tests stay fast and isolated.
• Start with: flat, so results are exact and repeatable.

11. One database per user or tenant

Desktop apps and services that keep each user’s data apart.

• Why SrvDB: each database is a separate directory; deleting a user means deleting their directory.
• Note: each directory can be open in one place at a time, so keep one Database object per user in a single process.

When to choose something else

SrvDB is deliberately small. Look elsewhere if you need:

• A shared database server used by many machines or services: SrvDB is a library, single-node, with no network access.
• Several processes writing to the same database: only one can open it at a time.
• Filtering by metadata inside the search (for example where year > 2020): planned for 0.4; today, filter the results after searching or keep one database per category.
• Billions of vectors on one machine: SrvDB has no published benchmarks at that scale yet.

By industry

See 100 projects in 20 industries, five levels each from beginner to frontier, every one with its own page: cybersecurity, AML, fiscal auditing, GRC, aerospace, mechanical and electrical engineering, quantum computing, energy, healthcare, legal, finance and more.

Live examples

Runnable, end-to-end examples and cheat sheets for these use cases are coming soon.