Retrieval troubleshooting

Diagnosing Aimogen Pro embeddings problems: nothing retrieved, wrong results, stale answers, dimension errors and unexpected cost.

Nothing is retrieved#

Symptoms. The AI answers as though it has no knowledge base.

Work through in order:

1. Is retrieval enabled for this context? Settings › Embeddings › Enable Embeddings For. For the chatbot this is Chatbot Shortcodes. This is the most common cause by a wide margin.
2. Does the namespace match? Content indexed into one namespace is invisible to a context configured for another. Check both, including the embeddings_namespace shortcode attribute.
3. Is anything indexed? Check AI Embeddings › List Added Embeddings.
4. Is the vector store reachable? Check System & Logs › Activity Logs for connection errors.
5. Is the score threshold too high? A high OpenAI Vector Store Minimum Score filters everything out.
6. Is the result count zero or one? Pinecone, Qdrant and Chroma default Number Of Results To Query to 1. Raise it to 3–5.

Retrieved content is irrelevant#

1. The embedding model was changed after indexing. This is the quiet killer — see below.
2. Chunks are too large. A whole long post as one chunk produces a vague vector that matches everything weakly.
3. The embedding template is poor. Include the title and taxonomy so chunks carry context. See Adding content.
4. The namespace is too broad. Searching everything returns the topically-nearest thing, not the correct thing.
5. No score threshold. Weak matches are being treated as relevant.

Answers are out of date#

The index reflects content as it was when indexed.

• Re-index edited posts.
• Delete entries for deleted posts — orphaned vectors still get retrieved.
• Enable auto-indexing for new posts.
• Re-index after any bulk edit.

Dimension mismatch errors#

Symptom. The vector store rejects vectors, mentioning dimensions.

The index was created with a different dimension than the embedding model produces. The plugin help text specifies dimension 1536, metric cosine for Pinecone, Qdrant and Chroma, matching text-embedding-3-small.

Fix: recreate the index at the correct dimension and re-index from source.

Everything got worse after changing the embedding model#

The chatbot ignores the knowledge base#

Almost always Chatbot Shortcodes not enabled under Settings › Embeddings › Enable Embeddings For. Indexing content and enabling retrieval for the chatbot are two separate steps.

Then check the namespace matches, and that the model is being told to use the reference information — see Retrieval.

Indexing is very slow#

Each document is an API call plus a store write. A thousand posts is a thousand of each.

• Index in batches rather than all at once.
• Maximum Number Of Unindexed Articles To Process limits each batch for duplicate detection indexing.
• If Optimize The %%post_content%% Shortcode Using AI is on, each document also costs a full AI rewrite. Turn it off for a large initial index.

Costs are higher than expected#

Indexing is cheap; retrieval injection is not. Every retrieved chunk is input tokens on every affected request.

• Reduce the result count.
• Raise the score threshold.
• Enable retrieval only for the contexts that need it, not all seventeen.

See Controlling cost.

Related posts are wrong#

[aimogen-related-posts] uses similarity. Tune with the threshold attribute:

[aimogen-related-posts limit="5" threshold="0.8"]

Raise it for fewer, closer matches. If nothing appears, the posts are not indexed.

Agent long-term memory is unavailable#

The option is disabled with an explanatory note when embeddings are not fully configured: you need an API key, an embedding model and a vector store. See Creating an agent.

Vector store connection fails#

Also confirm your server can reach the host — an outbound firewall will produce the same symptom.

Getting help#

Collect:

1. Which vector store and embedding model
2. The index dimension
3. Which contexts have retrieval enabled, and their namespaces
4. Errors from System & Logs › Activity Logs
5. System & Logs › System Info

Then open a support ticket.

Related#

• Vector stores
• Embedding models
• Retrieval and prompt injection