Modernize, refine, and standardize the notebook collection by sis0k0 · Pull Request #219 · mongodb-developer/GenAI-Showcase

sis0k0 · 2026-07-22T11:58:58Z

Overview

This PR is a broad maintenance and quality pass over the notebook collection, focused on notebooks/agents/. It normalizes install commands, cleans up cell outputs, adds "Open in Colab" buttons, standardizes environment-variable handling, migrates notebooks to the current Atlas vector index syntax, fixes dead links, and substantively refines several individual agent notebooks. 35 files changed.

Deeply refined notebooks

Substantive content and narrative rewrites, not just mechanical edits:

agent_fireworks_ai_langchain_mongodb.ipynb: Fully modernized the Fireworks + LangChain agent, with refreshed narrative, reworked setup, and updated code.
ai_agent_with_pydanticai_and_mongodb.ipynb: Refined the Pydantic AI + MongoDB agent (~2,000 lines reworked).
agentchat_RetrieveChat_mongodb.ipynb: Refined the AutoGen AgentChat / RetrieveChat walkthrough.
agentic_rag_factory_safety_assistant_with_langgraph_langchain_mongodb.ipynb: Compatibility fixes and deprecation-warning cleanup (~1,900 lines), plus an expanded intro and problem statement.
airbnb_agent_openai_llamaindex_mongodb.ipynb: Refined narrative and setup.

Standardized environment-variable handling

The five modernized notebooks now load secrets consistently:

Add python-dotenv to the install cell and call load_dotenv() so variables defined in a local .env file are read automatically.
Introduce a shared get_or_prompt_env(var_name, prompt_text) helper that returns the value from the environment (.env or already-set) when present, and otherwise falls back to a getpass prompt, raising a clear EnvironmentError if left empty.
Keys such as OPENAI_API_KEY, FIREWORKS_API_KEY, and MONGODB_URI are resolved through this helper instead of being hardcoded or prompted ad hoc, so the notebooks run non-interactively when a .env file is present and prompt gracefully when it is not.

Other standardized conventions

Applied consistently across the agent notebooks:

Install commands: Replaced !pip with %pip, then normalized all installs to %pip install -U -q for quiet, up-to-date installs.
Cell outputs: Cleared noisy pip-install outputs.
Open in Colab: Added launch buttons to all notebooks across agents/, rag/, evals/, and partners/.
Intro sections: Added consistent titles and problem statements to ~15 agent notebooks.

Modernization

Atlas vector index syntax: Migrated the remaining notebooks (advanced retrieval techniques + LangChain, multimodal Atlas vector store, POLM AI Stack) to the current syntax.

Housekeeping

Dead-link fixes: Patched broken links in langchain_parent_document_retrieval.ipynb, the multimodal Atlas notebook, and the Haystack cooking-agent notebook.
.gitignore: Added one entry (to keep local .env files out of version control).

ajosh0504 · 2026-08-18T23:46:18Z

Ran a code review agent on the PR and spot-checked its outputs and recommendations. Seemed sane to me. Here at it's findings: PR-219-review.md

…ll order ag2 1.0.2 dropped the retrievechat-mongodb extra and the autogen namespace, so the notebook failed on both %pip install and the subsequent autogen imports. Pin ag2<1.0 with the openai extra (needed by autogen.oai.client but not pulled in transitively), and move the install cell above the dotenv/pymongo/openai imports it was previously running after.

…niques notebook SearchIndexModel was migrated to new-style vector field syntax but never got type="vectorSearch", so Atlas rejects the index creation with "Attribute mappings missing." The surrounding try/except only printed the error and let execution continue, masking the failure until every later $vectorSearch stage broke with no obvious cause. Now the type is set and the exception propagates instead of being swallowed.

…sion section max_tokens=256 left no budget for gpt-oss-20b's reasoning output, so both agent invocations returned an empty string, the empty answer got written into MongoDB chat history, and the knowledge_base tool was never reached in the second turn. Raise max_tokens to 4096 and fail loudly instead of silently returning '' when a model call comes back empty. Also: retitle the "compression retriever" section to describe what it actually does (compression is exposed as an explicit tool, not a wrapping retriever) instead of building an unused variable; restore the gpt2 pin on PromptCompressor so it doesn't fall back to a ~13GB default model on CPU; drop the unreachable device_map except branch; correct the prose naming the wrong Fireworks model; pin the fast-moving langchain packages instead of resolving to latest. Stale committed outputs from the max_tokens=256 bug (leaked harmony markers, empty answers) are cleared rather than left in, since they can't be re-executed without live API credentials in this session.

…ynced docs MIN_FOREIGN_RESULT_SCORE was set to 0.0, making the score < threshold check unreachable for any non-negative Tavily score, so every foreign search result got embedded and written to working_memory regardless of relevance. Restore the 0.5 threshold. Also fix markdown/code desync left by an earlier reshuffle: move the agent demo call and its wrap-up back next to the agent definition, move the dataset-loading narrative to actually precede the dataset loading code, remove a verbatim duplicate agent-definition cell, and restore the missing "Step 3: Data Loading and Preparation" heading (the notebook was jumping from Step 2 straight to Step 4). Pin pydantic-ai instead of resolving to latest.

…scaffolding ANTHROPIC_API_KEY was required via get_or_prompt_env and raised OSError on an empty prompt, but the notebook actually runs on ChatOpenAI(model="gpt-5-mini") — ChatAnthropic is commented out. This blocked anyone without an Anthropic account from getting past that cell. Drop the requirement and the unused langchain_anthropic install; fix the invalid commented-out model id (claude-sonnet-4 -> claude-sonnet-4-0) and note that temperature=0 is silently ignored by langchain-openai for gpt-5-mini. Also remove leftover debug scaffolding: a final cell that made an extra billed LLM call to inspect graph output shape (with its committed output), and max_retries=1 which made the retry branch dead code. Pin the fast-moving langchain/langgraph packages instead of resolving to latest.

A prior dead-link pass mistook pandas' repr ellipsis truncation (".../access/manage-t...") for a broken link and rewrote it to a URL that doesn't match its own row label. Revert the committed outputs back to what the code actually produced rather than keeping the fabricated substitutions.

- Fix glued H1 headings in 5 notebooks where the heading was appended to the Colab badge markdown line with no trailing newline, so it rendered as literal text instead of a heading. - Remove the templated filler problem statement ("This notebook solves the problem of building and evaluating <filename> workflows using MongoDB-backed retrieval and agent orchestration.") generated from the filename in 11 notebooks, rather than leaving generic boilerplate in place of a real description.

_search_arxiv created a brand-new arxiv.Client() on every call, so its built-in "no more than one request every 3 seconds" throttling was reset each time and never actually applied between calls. Once the earlier max_tokens fix let the agent make several arXiv searches within one turn, the back-to-back requests weren't spaced out and arXiv started returning HTTP 429. Share a single Client instance across calls so its rate limiting works as intended.

sis0k0 · 2026-08-19T10:50:55Z

@ajosh0504 , thank you for the review! I addressed the comments in the latest commits.

ajosh0504 · 2026-08-19T15:53:56Z

@sis0k0 Reran the code review agent. Few more minor issues that result in failures when running the notebook outside of Colab:

Install nest_asyncio inai_agent_with_pydanticai_and_mongodb.ipynb
Install nest_asyncio in agentic_rag_factory_safety_assistant_with_langgraph_langchain_mongodb.ipynb
[Optional] Clean up doubled-flag pip lines in notebooks (Eg: -U -q -U langgraph, -U -q -Uq openai).

Everything else looks good!

Both notebooks call nest_asyncio.apply() but never installed the package, relying on it being preinstalled in Colab. Add it to the pip install cell so the notebooks run outside Colab too.

Several install cells had redundant/duplicated flags from a prior mechanical edit, e.g. "-U -q -U langgraph" and "-U -q -Uq openai". Dedupe them down to a single -U -q.

sis0k0 · 2026-08-20T05:46:20Z

@ajosh0504 , good catch! Cleaned up the doubled install flags and add the missing nest_asyncio installs.

sis0k0 added 15 commits July 2, 2026 17:48

Apply phase 2 notebook crawlability updates

b6ba330

Clear pip install outputs from agent notebooks

f754ff4

Replace !pip with %pip in agent notebooks

f5e61f4

Add Open in Colab buttons to all notebooks

433dab8

Patch dead notebook links and add link audit report

a3bb620

Normalize agent notebook installs to %pip -U -q

3b6506e

Add intro title and problem statement to agent notebooks

8be75ae

Migrate remaining notebooks to current Atlas vector index syntax

728a926

Modernize the Fireworks+Langchain Agent notebook

c05e85e

Cosmetic changes

91ac8bd

Refine agent notebook narrative and env handling

3d3540a

Refine AgentChat notebook

ff0e989

fix notebook compatibility updates and warning cleanup

f2a53cf

Refine AI Agent with Pydantic and MongoDB notebook

717e216

Remove extraneous artifacts

e1ff6b2

sis0k0 requested a review from a team as a code owner July 22, 2026 11:58

Apply pre-commit fixes (ruff-format, ruff, widget state)

040b579

sis0k0 added 8 commits August 19, 2026 09:39

sis0k0 added 2 commits August 20, 2026 08:43

Install nest_asyncio explicitly where it's used but not installed

65fe6a5

Both notebooks call nest_asyncio.apply() but never installed the package, relying on it being preinstalled in Colab. Add it to the pip install cell so the notebooks run outside Colab too.

Clean up doubled pip install flags in agent notebooks

e79c029

Several install cells had redundant/duplicated flags from a prior mechanical edit, e.g. "-U -q -U langgraph" and "-U -q -Uq openai". Dedupe them down to a single -U -q.

Navigation Menu

Uh oh!

Modernize, refine, and standardize the notebook collection - #219

Modernize, refine, and standardize the notebook collection#219
sis0k0 wants to merge 26 commits into
mainfrom
update-agents-notebooks

sis0k0 commented Jul 22, 2026

Uh oh!

ajosh0504 commented Aug 18, 2026 •

edited

Loading

Uh oh!

sis0k0 commented Aug 19, 2026

Uh oh!

ajosh0504 commented Aug 19, 2026

Uh oh!

sis0k0 commented Aug 20, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

sis0k0 commented Jul 22, 2026

Overview

Deeply refined notebooks

Standardized environment-variable handling

Other standardized conventions

Modernization

Housekeeping

Uh oh!

ajosh0504 commented Aug 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

sis0k0 commented Aug 19, 2026

Uh oh!

ajosh0504 commented Aug 19, 2026

Uh oh!

sis0k0 commented Aug 20, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

ajosh0504 commented Aug 18, 2026 •

edited

Loading