Agently Knowledge Base And RAG

This skill covers Agently's current knowledge-base and RAG path built around agently.integrations.chromadb. It focuses on embedding-agent-backed indexing, Chroma collection usage, retrieval, and the common pattern of injecting retrieval results into a normal request. It does not attempt to document generic vector-database strategy outside Agently's current Chroma integration surface.

Prerequisite: Agently >= 4.0.8.5.

Scope

Use this skill for:

• ChromaCollection
• embedding-agent-backed indexing
• add(...), query(...), and query_embeddings(...)
• query results injected through info(...)
• lower-level ChromaData or ChromaEmbeddingFunction
• process-level reuse of one knowledge base across many requests

Do not use this skill for:

• plain embeddings requests without a knowledge base
• generic vector-database choices outside the current Chroma integration
• TriggerFlow workflow orchestration as the main problem
• service architecture as the main problem

Workflow

1. Start with references/chromacollection-basics.md when building or querying a collection.
2. Read references/retrieval-to-answer.md when turning retrieval results into a normal answer flow.
3. Read references/process-lifecycle.md when the real-world issue is build-once reuse or long-lived service behavior.
4. Read references/chromadata-and-embedding-function.md when integrating with a lower-level Chroma client or precomputed data objects.
5. If the task is only about embedding vectors, switch to agently-embeddings.
6. If the task becomes flow orchestration or long-running agent loops, switch to agently-triggerflow-playbook.
7. If behavior still looks wrong, use references/troubleshooting.md.

Core Mental Model

Agently's current knowledge-base pattern is:

1. build or connect a Chroma collection
2. use an embedding agent to index documents
3. query the collection with a user question
4. inject retrieval results into a normal agent request
5. answer with retrieval context

The retrieval layer and the answer layer are separate on purpose.

Selection Rules

• just need embeddings -> agently-embeddings
• need a retrievable document collection -> ChromaCollection
• need lower-level Chroma integration objects -> ChromaData or ChromaEmbeddingFunction
• retrieval results should guide a normal answer request -> query then inject through info(...)
• knowledge base should be reused across turns or requests -> build once and keep the collection outside the request loop

Important Boundaries

• this skill documents Agently's current Chroma-backed KB/RAG path, not generic vector-database architecture
• retrieval should stay separate from final answering so the answer request can still use normal Agently prompt and output control
• process-level reuse matters in real services; avoid rebuilding the whole collection on every turn when the corpus is stable

References

• references/source-map.md
• references/chromacollection-basics.md
• references/retrieval-to-answer.md
• references/process-lifecycle.md
• references/chromadata-and-embedding-function.md
• references/troubleshooting.md