mongosemantic

Connection Collections Search Query Dashboard Visualize MCP Guide
…

How to use
  1. Inspect opens field-level statistics and lets you configure semantic search.
  2. Migrate model appears on configured rows — switch embedding models with near-zero downtime.
  3. The Model column shows the active embedding model and storage mode (shadow or inline).

How to use
  1. The table scores every field for semantic-search suitability — look for Great or Good.
  2. Click Configure semantic search to pick which fields to embed.
  3. Scroll the sample list at the bottom and click any row to inspect a full document.
Bands: Great = long, varied text. Good = usable. Usable = short/sparse, combine with another field. Not recommended = IDs, numerics, dates.

Field analysis

Configure semantic search

Sample documents

Click any row to inspect the full document.

Document 

    


    

      

      

      

        How to use
        
    
          
  1. Fields: comma-separated paths (e.g. title, body). Each field gets its own embedding column.
    2. 
          
  2. Mode: shadow writes embeddings to {collection}_embeddings. inline writes to _msem.{field} on each source doc.
    3. 
          
  3. Chunking (shadow only): splits long text so search returns the best paragraph, not the best doc.
    4. 
          
  4. Model: local-fast is free and runs on your CPU. OpenAI / Ollama need their respective env vars.
    5. 
        

        Tip: for most cases start with shadow + local-fast. You can migrate models later without downtime.
      

      

      

    


    

      

      

      

        What you're seeing
        
    
          
  1. Enqueueing writes one embed job per (document, field, chunk). It's fast and synchronous.
    2. 
          
  2. The embedded worker drains the queue in the background — you can leave this page, the work continues.
    3. 
          
  3. This page polls every 1.5 s so you can watch progress, throughput, and any failures live.
    4. 
        

      

      

        

          

          
        

        
        

      

      

        

          Completed
          —
        

        

          In flight
          —
        

        

          Pending
          —
        

        

          Failed
          —
        

        

          Worker
          
            
            —
          
        

      

      

        
Per-field progress

        

      

      

        
Failed jobs 

        
These jobs hit max attempts. Inspect the error and rerun retry from CLI or the dashboard.

        

      

      

        
Recent activity

        
Most recent completed and failed jobs. Newest on top.

        
    
      
    
    
    

    
    
      
    Search
    
      
    
      
    
        How to use
        
      
          
    1. Type a natural-language query — describe what you want, don't pick keywords.
      2. 
          
    2. Leave the collection dropdown on All configured collections to fan out, or pick one to scope.
      3. 
          
    3. Press Enter or click Search.
      4. 
          
    4. Toggle Hybrid for semantic + keyword (shadow mode) — works on every topology.
      5. 
        
    
        Try: spies blackmail cold war Berlin, lonely astronaut stranded, or tragic love story during a war.
      
    
      
    
        
          
          
        
        
        
      
    
      
    
        
        
        
        
        
        
      
    
      
    
        Export current results:
        
        
        
      
    
      
    
      
    
      
    
      
      
    
      

    
      
      
      
      
      
      
      
        How to use
        
        
          
      1. Pick a collection (any collection — not just configured ones).
        2. 
          
      2. Pick a quick example or write a pipeline as a JSON array of stages.
        3. 
          
      3. Click Run. Hard limits: 10-second timeout, 100-row result cap.
        4. 
          
      4. Blocked at parse time: $out, $merge, $function, $accumulator, $where, $jsonSchema.
        5. 
        
      
      
      
      
      
        
        
      
      
      
      
      
      
        
        
        
          
          
        
        
          
          
        
      
      
      
      
        
      
        
      
      
      
    
      

    
      
      
      Overview
      
      
      Topology, configured collections, and queue health at a glance.
      
      
      
        How to use
        
        
          
      1. Cards summarize the database state. Pending = jobs the worker hasn't processed yet.
        2. 
          
      2. Workers shows every worker that's heartbeated recently (running < 30s old, stale < 5min, dead otherwise).
        3. 
          
      3. Recent failures shows the last 10 failed embed jobs with their error message.
        4. 
          
      4. Retry failed jobs resets all failed jobs back to pending.
        5. 
        
      
        To see Workers populate: run mongosemantic worker from a terminal.
      
      
      
      
      
      
      
      Indexing activity auto-refreshing every 3s
      
      
      
      
      Workers
      
      
      
      
      Recent failures
      
      
      
    
      

    
      
      
      Explore embeddings
      
      
      2D PCA projection of sampled embeddings. Points close together mean similar by meaning.
      
      
      
        How to use
        
        
          
      1. Pick a collection. If it has multiple configured fields, pick which one to plot.
        2. 
          
      2. Each dot is one embedding (or one chunk, for chunked collections).
        3. 
          
      3. Hover any dot to see its source text snippet.
        4. 
          
      4. Look for visual clusters — they correspond to semantically similar groups of documents.
        5. 
        
      
        Tip: increase the Sample dropdown for richer clusters — 5k is the sweet spot, 50k loads every embedding (slower but shows the full structure).
      
      
      
      
        
        
        
        
        
      
      
      
      
      
      
        
      
          
          
      
        
      
        
      
          
      Clusters
      
          
      Click a cluster to highlight. Click again to clear.
      
          
        
        
        
      
        
      
        
    
        

    
    
        
      
        
        
        Migrate model
        
        
        
        
        
        
        
        
        
        
        
          
          
        
        
      
        
    
        

    
        
      
        
      
        
      
        
        How to use
        
          
          
        1. Run the integrate claude command in a terminal — it writes mongosemantic into Claude Desktop's config file.
          2. 
          
        2. Quit and reopen Claude Desktop. The mongosemantic tools appear in the slash menu of every new chat.
          3. 
          
        3. Ask Claude something like "use mongosemantic to find movies about isolated astronauts" — it picks the right tool automatically.
          4. 
        
        
        Other clients: Cursor or anything that speaks MCP over SSE can connect with mongosemantic serve --transport sse.
      
        

      
        Claude Desktop
        
      
        
      
        
      
        

      
        Other MCP clients (Cursor, SSE)
        
      
        
      
        

      
        
      
        
    
        

    
        
      
        Guide
        
      
        A 5-minute tour of everything mongosemantic does. Each section maps to a page in the nav above.