All models tested with ALLM(AnythingLLM) with LM as server

They work more or less (sometimes the results are more truthful if the “chat with document only” option is used)

some models can not hande large TXT files (maybe only 200pages - hints below)

My short impression:

nomic-embed-text
mxbai-embed-large
mug-b-1.6
Ger-RAG-BGE-M3 (german)

Working well, all other its up to you!

Short hints for using: Set your (Max Tokens)context-lenght 16000t main-model, set your embedder-model (Max Embedding Chunk Length) 1024t,set (Max Context Snippets) usual 14, but in ALLM its cutting all in 1024 character parts, so aprox two times or bit more ~30!

-> Ok what that mean!

You can receive 14-snippets a 1024t (14336t) from your document ~10000words and 1600t left for the answer ~1000words (2 pages)

You can play and set for your needs, eg 8-snippets a 2048t, or 28-snippets a 512t ...

8000t (~6000words) ~0.8GB VRAM usage /
16000t (~12000words) ~1.5GB VRAM usage /
32000t (~24000words) ~3GB VRAM usage

...

How embedding and search works for now

You have a txt/pdf file maybe 90000words(~300pages). You ask the model lets say "what is described in chapter called XYZ in relation to person ZYX". Now it searches for keywords or similar semantic terms in the document. if it has found them, lets say word and meaning around “XYZ and ZYX” , now a piece of text 1024token around this word “XYZ/ZYX” is cut out at this point. This text snippet is then used for your answer. If, for example, the word “XYZ” occurs 100 times in one file, not all 100 are found (usually only 4 to 32 snippet are used)

A question for "summary of the document" is most time not useful, if the document has an introduction or summaries its searching there if you have luck.

If the documents small like 10-20 Pages, its better you copy the whole text inside the prompt.

...

My impression is that at the moment only the first results of the embedder is used (dont ask what happend if you use more than one document at once) so it should be some kind of ranking.

If the embedder has found 100 snippets, only the snippets with the highest rank should be sent to the model.... but seems thats not happend

...

on discord (sevenof9) ...

...

(ALL licenses and terms of use go to original author)

...

avemio/German-RAG-BGE-M3-MERGED-x-SNOWFLAKE-ARCTIC-HESSIAN-AI (German, English) - 600pages and more
maidalun1020/bce-embedding-base_v1 (English and Chinese) - only ~200pages
maidalun1020/bce-reranker-base_v1 (English, Chinese, Japanese and Korean) - only ~200pages
BAAI/bge-reranker-v2-m3 (English and Chinese) - 600pages and more
BAAI/bge-reranker-v2-gemma (English and Chinese) - dont work
BAAI/bge-m3 (English and Chinese) - 600pages and more
avsolatorio/GIST-large-Embedding-v0 (English)- ~300pages
ibm-granite/granite-embedding-278m-multilingual (English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese) - only ~200pages
Labib11/MUG-B-1.6 (?) - ~300pages
mixedbread-ai/mxbai-embed-large-v1 (multi) - ~300pages
nomic-ai/nomic-embed-text-v1.5 (English, multi) - 600pages and more
Snowflake/snowflake-arctic-embed-l-v2.0 (English, multi) - 600pages and more
intfloat/multilingual-e5-large-instruct (100 languages) - only ~200pages
T-Systems-onsite/german-roberta-sentence-transformer-v2 - ~300pages