Knowledge map on a local model. 23,457 texts processed on our server
Authorship and approval
Published and updated
We moved text vector calculations from Gemini to a locally run EmbeddingGemma. We reprocessed the entire archive, building the map for 23,457 texts in 72 seconds. Visually, the new map better shows transitions between topics, but the tag-based comparison does not settle which one is better. The neighbor test on manually labeled texts is still ahead of us.
Highlights
- EmbeddingGemma 2 locally processed 23,163 archival texts in 10.2 hours, and another 294 in about 8 minutes.
- The map for 23,457 texts was generated from ready vectors in 72 seconds and contains 38,994 edges connecting neighboring texts.
- Visually, the EmbeddingGemma map creates continuous thematic areas, but the tag-based comparison result is skewed in favor of Gemini.
- A shared vector table already serves the map and the Fiszka classifier; subsequent services will transition to it after tests.
What is a knowledge map
We collect texts about AI daily: arXiv papers, company blog posts, and news. We also include our own posts and write-ups. The knowledge map shows how similar these texts are to each other.
We turn each text into an embedding, which is a list of numbers describing its content. Texts with similar themes receive similar numbers. On the page, we reduce this multidimensional data to three dimensions using UMAP. Each point corresponds to a single text, and closely positioned points usually concern similar topics. An edge connects a text to its two nearest neighbors.
What we changed
Previously, we calculated vectors for the map using Gemini Embedding (gemini-embedding-2). The model received 1536 numbers per text via the Google API in the cloud. We replaced it with EmbeddingGemma 2 (embeddinggemma-2:270m) — an open-weights model from Google, run locally via Ollama on our server. It has about 271 million parameters, and the model file weighs about 303 MB. Now each text receives 768 numbers.
What goes into the model has also changed. Gemini received a title, a summary written by Gemini, and text tag names, for example "agents, language models". The whole thing was a few hundred characters long. EmbeddingGemma receives the title and up to 4000 characters of actual content, without tags, formatted as a model card document: title: … | text: ….
Element | Previously | Now |
|---|---|---|
Model | Gemini Embedding (gemini-embedding-2) | EmbeddingGemma 2 (embeddinggemma-2:270m) |
Computation place | Google API in the cloud | Locally via Ollama on our server |
Vector dimension | 1536 numbers per text | 768 numbers per text |
Input data | Title, Gemini summary and tags | Title and up to 4000 characters of content, without tags |
We also organized how vectors are stored. Previously, the same texts had three sets: Gemini for the map, separate local chat vectors, and a third set for the Fiszka classifier. Now the map and Fiszka use a single vector table. This same table now feeds the Fiszka 0.3 topic classifier, which we wrote about on November 6.
One lab loop collects texts and calculates vectors for them in the morning, during a window when the server has no other work. We separated these two tasks, instead of combining text collection with processing them. Chat, searching for similar texts, and selecting news for posts will transition to the shared table one by one.
How long the processing took
Logs show that processing the archive covering 23,163 texts took 10.2 hours, between November 6 and November 7. On November 8, we calculated vectors for 294 texts that arrived in the meantime. This took about 8 minutes, or about 1.6 seconds per text when processing batches of 16.
Once the vectors were ready, building the map for 23,457 texts took 72 seconds. Fetching data from the database took from 1 to 3 seconds, finding nearest neighbors 11 seconds, and reducing 768 dimensions to three in UMAP — 59 seconds. The result was 38,994 edges.
The previous process was limited by the free tier of Gemini Embedding: ours had a ceiling of about 1000 texts per day. We measured the limit on August 21; in practice, the loop processed 40 texts per hour. The local model has no such daily API limit. We are limited by the time available on the server.
Is the new map better?
We compared, among other things, what percentage of edges connect two texts from the same thematic group. The Gemini map reached 69.8%, and the EmbeddingGemma map 50.6%. This number alone speaks in favor of Gemini, but does not settle which map better reflects the content of the texts.
We found the reason in the input data. The Gemini vector was created from text containing tag names. Our metric checks the compliance of neighbors with the same tags. Gemini therefore had words in its vector whose presence the metric later rewarded. This result tells whether tags match tags, rather than whether the map shows content similarity well. We realized this only after the calculations.
Visually, the Gemini map was scattered across corners, with separate islands for tag combinations. The EmbeddingGemma map looks like a single continuous area with a dense center and transitions between topics. For example, topics concerning agents transition into language models and tools. Distant topics, such as papers on proteins and biology, lie mostly separate. This is an observation from viewing the map, not the result of a decisive quality test.
A fairer test is still ahead of us: we will check neighbors for manually labeled texts, with tags that do not come from any model. We will announce it in a subsequent post, but we do not have the result yet.
What we learned and what's next
We learned that a metric can look solid and still answer a different question than the one we want to ask. When comparing models, you need to check what the input for each of them was calculated from. Tacking tags onto text before calculating the vector can turn a content map into a label map.
We also saw the practical consequence of maintaining multiple vector sets for the same texts: they can drift apart. On November 8, the classifier page failed to refresh precisely because of a gap between sets. Hence the decision for a single table and a single process that populates it.
Chat will transition to the same table after the response quality test. Later we will move the search for similar texts, selection of news for posts, and notes. Finally, we will disable the Gemini vectors. Testing the map on manually labeled texts remains a separate task.
Final remarks
There is no team behind the lab: experiments are prepared by a single person together with AI. The project is a hobby, done for learning and fun.
Frequently asked questions
What is an embedding?
An embedding is a list of numbers describing text content. Each text receives such a vector, and texts with similar content have similar numbers. In the knowledge map, vectors are used to position texts and find their nearest neighbors.
Why a local model instead of Gemini?
A local model provides a shared vector table for the map, the classifier, and subsequent services such as chat. It is not limited by a daily API limit, and data does not leave for the cloud during vector calculation.
Is the new map better?
Visually, the EmbeddingGemma map better shows continuous areas and transitions between topics. Tag-based comparison favors Gemini, but it is skewed: tags were present in Gemini's input data. The neighbor test on manually labeled texts has not been conducted yet.
How long does processing take?
Processing 23,163 archival texts took 10.2 hours. Building the map from ready vectors for 23,457 texts took 72 seconds.
Sources
- EmbeddingGemma, Google documentation
- EmbeddingGemma 2 in the Ollama library
- Gemini API: embeddings
- UMAP, documentation
How to cite
aipoc-lab. "Mapa wiedzy na lokalnym modelu. 23 457 tekstów przeliczonych na naszym serwerze". November 8, 2026. https://aipoclab.creativeaidrivenpm.com/posts/mapa-wiedzy-na-lokalnym-modelu-23-457-tekstow-przeliczonych-29065
BibTeX: @misc{aipoclab2026_29065, title = {Mapa wiedzy na lokalnym modelu. 23 457 tekstów przeliczonych na naszym serwerze}, author = {aipoc-lab}, year = {2026}, month = nov, howpublished = {https://aipoclab.creativeaidrivenpm.com/posts/mapa-wiedzy-na-lokalnym-modelu-23-457-tekstow-przeliczonych-29065}, note = {Blog post}}
--- Text prepared by the aipoc-lab (AI) laboratory based on its own lab notes with measurements and conclusions. Version before human editing.
Provenance manifest of this post (JSON) — which signals it was built from, what decisions the loops made, and when a human approved it.
Related posts
Fiszka 0.3 is public. The model, model card, and a week in production alongside Gemini
Fiszka 0.3 weights, model card, and code are now public, showing a 75% topic agreement with Gemini over a one-week production test.
AI Today: llama.cpp 1.0.0 and Shield
Discover the latest AI updates including llama.cpp 1.0.0, Shield for IT portals, the XYBench benchmark, Isaac Sim, and Dyania Health.
AI Today: ChatGPT Wallet and DeepSeek API
Discover how ChatGPT Wallet, DeepSeek API, Google WeatherNext 3, and Gemini updates shape AI applications in mid-November 2026.
AI Today: DeepSeek V4.1-Flash and Kimi K3.1 Pro
Discover the latest updates on DeepSeek V4.1-Flash, Kimi K3.1 Pro, Muse voice chats, Apple CarPlay integration, and Gemini Omni 1.2 Flash.