Home / Features / Knowledge and memory

Knowledge bases, memory, and RAG on osFoundry

Auto-extracted entities, knowledge graphs, custom retrieval pipelines, and reranked RAG — all configurable in osStudio.

osFoundry is a knowledge platform built around AI: every note, document, and uploaded file is auto-chunked, embedded, and indexed. A workspace knowledge graph extracts entities and relationships across everything you’ve worked on. Custom retrieval pipelines (with reranking, MMR diversity, and multi-stage stages) are configurable per chat path in osStudio — no source-code changes needed.

Quick answer

Auto-chunking pipeline indexes notes, KBs, and files as they’re created.
Knowledge graph extracts entities + relationships continuously, queryable via the Graph screen.
Custom RAG pipelines — define stages (retrieve, filter, rerank, MMR) per Path config.
Voyage embeddings + reranking through a managed proxy (no separate API key).

What it is

A knowledge base is only useful if the right passage surfaces at the right moment. osFoundry treats that as a pipeline problem: every note, document, and uploaded file is chunked, embedded, and indexed as it is created, so retrieval is ready the moment content exists — there is no separate “ingest” step to remember and no vector database to operate.

Retrieval is hybrid and configurable. Under the hood a query runs through embedding search and BM25 keyword search, then optional reranking and MMR diversity stages; each Path in osStudio can define its own stages, models, and thresholds. This matters because stuffing whole documents into a prompt hits context-window limits and degrades answer quality — retrieving only the relevant chunks keeps the context focused.

Two layers accumulate on top. The workspace knowledge graph extracts entities and relationships from every chunk continuously, queryable from the Graph screen. And when long conversations compact, memory extraction distills them into durable facts — so decisions made in chat three months ago remain findable, not buried in scrollback.

Key capabilities

Notes with rich Markdown, version history, and three-way merge.
Knowledge bases — versioned, multi-author, exportable.
Workspace knowledge graph with entity extraction on every chunk.
Memory extraction on chat compaction — long conversations distilled into facts.
Configurable retrieval pipelines (Voyage embed → tantivy BM25 → reranker → MMR).
Workspace-wide search across notes, KBs, chats, and code.

How to do it in osFoundry

Add sources — Write notes, create knowledge bases, or upload files. Chunking, embedding, and indexing happen automatically on creation.
Ask with grounding — Chat normally — the retrieval pipeline attached to your Path pulls the relevant chunks into context and the model answers against them.
Tune retrieval when needed — In osStudio, adjust the pipeline per Path: swap embedding models, add a reranker, tighten thresholds, or add MMR diversity — no source-code changes.

Use cases

Support teams: Load product docs and past tickets into knowledge bases; agents answer new tickets grounded in the actual documentation.
Researchers and writers: Digest long reports without pasting them into chat — retrieval brings in the relevant sections, avoiding context-window overflow.
Teams with long-running projects: Decisions distilled from compacted chats plus the knowledge graph answer “when did we decide this and why” in seconds.

Frequently asked questions

Does osFoundry use my data to train its models?

No. We don’t train on your data. Embeddings stay in your workspace; retrieval queries are scoped to your workspace; nothing crosses tenant boundaries.

What embedding model does osFoundry use?

Managed Voyage AI by default (no separate API key). You can swap to a local embedding model or BYOK to OpenAI/Cohere via the embedding pipeline config in osStudio.

Can I customise the RAG pipeline per chat?

Yes. Each "Path" in osStudio has its own retrieval pipeline — different stages, models, thresholds, rerankers per use case.

Is the knowledge graph generated automatically?

Yes — entities and relationships are extracted from every note, KB doc, and chat. Browse and edit it in the Graph screen.

Do I need to run a vector database?

No. Chunking, embedding, BM25 indexing, and reranking are built into the platform — there is no separate database to deploy or operate.

How is this different from pasting a document into the chat?

Pasting consumes the context window and degrades on long inputs. Retrieval selects only the chunks relevant to the question, which keeps context focused and works no matter how large the corpus grows.