MongoDB Search Playground Chatbot Demo Builder for RAG Applications

mongodb · Tech · 2024 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: RAG prototyping and vector-search developer experience

Developers want to understand and prototype RAG chatbots over their own data, but setting up Atlas accounts, clusters, collections, embeddings, vector indexes, chunking, and query inspection can slow early experimentation before teams know whether MongoDB Vector Search fits their use case.

Constraints:

• The builder is a public demo and explicitly warns users not to upload sensitive data.
• The environment is non-persistent except for snapshot URLs that last 30 days.
• Imported data is treated as a single unified knowledge source rather than multiple collections.
• The vector search index is preconfigured and the generated query cannot be edited directly.
• The demo is text-only, caps file and character sizes, and does not allow custom embeddings or API credentials.
• Question-answer pairs are independent and do not rely on previous interactions.

Architecture

MongoDB's Chatbot Demo Builder is a hosted Search Playground tool for trying RAG with MongoDB Vector Search. A user uploads a PDF, pastes text, or loads sample data, configures chunking and an embedding model, and then asks independent questions over the generated vector index. The backend uses Voyage AI embedding models, a preconfigured vector search index, and a MongoDB aggregation pipeline with $vectorSearch and $project. The UI exposes the generated search query, retrieved/scored documents, numCandidates, retrieval limit, data source, index definition, search query, LLM and prompt, snapshot sharing, and starter code.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  UI["Search Playground UI"] --> Import["Data import"]
  Import --> Chunking["Chunking configuration"]
  Chunking --> Embeddings["Embedding generation"]
  Embeddings --> Index["Vector search index"]
  UI --> Questions["Question interface"]
  Questions --> Pipeline["Aggregation pipeline"]
  Pipeline --> Index
  Pipeline --> LLM["LLM and prompt"]
  LLM --> Panels["Inspection panels"]
  Index --> Panels
  UI --> Panels
  UI --> Snapshot["Snapshot and starter code"]

Components

Data flow

The user imports a PDF, pasted text, or sample data into the Chatbot Demo Builder. The tool treats that import as one knowledge source, chunks text using recursive chunking or fixed token count with overlap, generates embeddings using a selected Voyage AI model, and stores searchable vector embeddings behind a preconfigured index. For each independent user question, the builder generates a MongoDB Vector Search query, runs $vectorSearch, projects the results, shows retrieved documents and scores, and sends context to an LLM and prompt for answer generation. Users can inspect settings and export a snapshot URL or starter code.

Why this architecture

MongoDB chooses a constrained playground architecture to lower RAG experimentation friction. Developers can use their own documents without creating an Atlas account or cluster, inspect the generated vector-search machinery, and try core retrieval settings quickly. The tradeoff is intentional: the environment is limited, non-persistent, and not equivalent to production performance, but it exposes enough of the production-shaped pipeline to teach chunking, embeddings, retrieval, scoring, prompts, and starter-code paths.

Models used: voyage-3-large · voyage-finance-2 · voyage-law-2 · Large Language Model shown in LLM & Prompt panel · MongoDB Vector Search

Scale

• PDF uploads are capped at 100 MB.
  “Upload a PDF file from your local device with a maximum size of 100 MB.”article, stated
• Pasted text and extracted PDF text are capped at 100,000 characters.
  “Copy and paste text up to 100,000 characters.”article, stated

  “The total character count cannot exceed 100,000 characters.”article, stated
• Chunk size can range from 40 to 1,500 tokens.
  “Minimum: 40 tokens”article, stated

  “Maximum: 1500 tokens”article, stated
• Chunk overlap can range from 0 to 750 tokens and must be at most half the chunk size.
  3 supporting quotes

  “The overlap size must be at most half of the chunk size.”article, stated

  “Minimum: 0 tokens”article, stated

  “Maximum: 750 tokens”article, stated
• Snapshot URLs persist for 30 days.
  “Use the Share button to generate a snapshot URL that persists for 30 days.”article, stated

Alternatives considered (roads not taken)

Require an Atlas account, cluster, and collection before trying RAG

Rejected because: The demo explicitly removes that setup requirement so users can ask questions over their own data immediately.

Let users edit the generated vector-search query directly

Rejected because: The builder generates the query from retrieval settings and does not allow direct code-editor edits in the demo environment.

Persist full demo environments and chat history by default

Rejected because: The playground is non-persistent; only data configuration and retrieval settings can be shared through a temporary snapshot.

Support multimodal PDF understanding

Rejected because: The builder only supports text-based embeddings and cannot process images inside PDFs.

Patterns

rag-retrieval (retrieval)

The builder demonstrates RAG by chunking imported data, generating embeddings, retrieving documents with MongoDB Vector Search, and using retrieved context to answer questions.

context-engineering (retrieval)

The playground exposes chunking strategy, chunk size, overlap, retrieval limit, and numCandidates so developers can tune what context reaches the model.

safety-policy-enforcement (safety)

The public demo constrains data, model customization, persistence, and sensitive-data use to keep the sandbox bounded.

Tradeoffs

Offer a no-infrastructure demo environment.

Chose Search Playground without Atlas account, cluster, or collection over Requiring full production setup before experimentation.

Rationale: The no-account path reduces initial RAG experimentation friction but necessarily limits persistence and customization.

Use preconfigured embeddings and index settings.

Chose Voyage AI model options and a non-editable vector search index over Bring-your-own embeddings, API credentials, and fully editable index/query code.

Rationale: Preconfiguration makes the playground easier to use, while limiting production-like customization.

Expose both recursive and fixed-token chunking.

Chose User-configurable chunking strategy, size, and overlap over A single hidden chunking policy.

Rationale: Different documents need different tradeoffs between natural structure preservation and predictable chunk sizes.

Keep Q&A independent rather than conversationally stateful.

Chose Independent question-answer pairs over Persistent chat history and shared conversations.

Rationale: The demo keeps sharing and evaluation centered on retrieval configuration rather than chat memory.

Gotchas (what bit them)

The demo is not suitable for sensitive data.

Root cause: It is a public demo and MongoDB logs workload data for monitoring and troubleshooting.

Resolution: Warn users not to upload sensitive data and verify generated answers before use.

The playground cannot model multi-collection RAG systems.

Root cause: Imported PDF or pasted text is processed as one unified knowledge source.

Resolution: Use the builder for initial exploration, then move to a production implementation when separate collections are needed.

The demo does not persist a full working environment or chat history.

Root cause: The playground is designed for temporary demonstrations, not durable production workspaces.

Resolution: Use the Share button to persist data configuration and retrieval settings for 30 days, and use starter code for durable implementations.

Image-heavy PDFs will produce incomplete answers.

Root cause: Only text-based embeddings are supported and images inside PDFs are ignored.

Resolution: Use text-only documents in the builder or implement a production multimodal pipeline outside the demo.

Evaluation

The builder exposes retrieval settings and outputs so users can manually inspect query syntax, retrieved documents, scores, numCandidates, retrieval limit, index definition, prompt, and LLM. It does not present a formal automated benchmark.

Metrics: Number of candidate matches reviewed · Number of chunks returned · Retrieved document scores · Generated Vector Search query syntax · Chunk size and overlap · Production-performance caveat

Outcomes

• MongoDB provides a low-friction RAG chatbot playground for trying vector search over user-provided data.
• Developers can inspect generated query syntax, retrieved documents, scores, index definition, data source, LLM, and prompt.
• Demo configurations and retrieval settings can be shared through 30-day snapshot URLs.
• Developers can use Get Code to access starter code for building a similar chatbot.

This is a developer demo environment rather than a production deployment. It supports 100 MB PDFs, 100,000 characters, chunk sizes from 40 to 1,500 tokens, overlap from 0 to 750 tokens, and 30-day snapshots.