Announcing the Rasa Developer Edition

Announcements
21

Hi! My rather limited exposure to Rasa OpenSource gave me some food for thoughts:

  1. RASA seems to be a very powerful platform
  2. There’s a marked difference between the OpenSource and Pro platforms, whereas OpenSource support is a) lame b) buggy c) version-outdated - it’s a killer! d) hardly expandable feature-wise, hence not reproducible. Try to add another story/entity/slot/action - forget about it! e) very clumsy on the forum side - what to ask and how to get help f) fuzzy and shallow in documentation … Isn’t it enough?!
  3. It looks very much like hooking folks on the Pro, especially with CALM being icing on the cake is the ‘hidden’ goal
  4. Asking ChatGpt for examples leads to endless fights with the system of Rasa to get things working - forget about quick prototyping, without shelling out 35k
  5. To me Rasa originally was a go-to platform to have proprietary buildup - not loosing the hope so far. Thanks a lot! Michael Smushkovich
22

Hello @amn41 its been great exploring rasa CALM so far.

although i notice that integration with gemini and hugging face has been really difficult. The instruction in the documentation doesnt seem to do justice as OPENAI_API_KEY error keep popping up even thought the config.yml was configured to use huggingface_llm.

google-generativeaicould not been installed has its not cmpatible with rasa-pro >= 3.10.X.

finally there no documentation on how to integrate rasa with qwen ai from aibaba cloud as well as deepseek r1.

23

thanks @kenfelix ! Can you share your config.yml and the error message you are getting? Would love to figure this out.

24

Hi, i have able to solve the issue. For using Gemini AI models, i did’nt need to install google-generativeai.

It seemed that it EnterpriseSearchPolicy and IntentlessPolicy uses Openai embeddings by default. Adding a different embedding solved the issue.

recipe: default.v1
language: en
pipeline:
  - name: SingleStepLLMCommandGenerator
    llm:
      model_group: gemini_llm
    flow_retrieval:
      active: false

policies:
  - name: FlowPolicy
  # - name: IntentlessPolicy
  #   llm:
  #     model_group: rasa_command_generation_model
  #   embeddings:
  #     model_group: huggingface_embedding_model
  - name: EnterpriseSearchPolicy
    llm:
      model_group: gemini_llm
    embeddings:
      model_group: huggingface_embedding_model

assistant_id: 20250217-101955-large-loop
25

However, I keep getting this feedback while training the model:

Provider List: https://docs.litellm.ai/docs/providers


Provider List: https://docs.litellm.ai/docs/providers

2025-02-18 06:31:35 INFO     faiss.loader  - Loading faiss with AVX512 support.

Provider List: https://docs.litellm.ai/docs/providers

2025-02-18 06:31:35 INFO     faiss.loader  - Could not load library with AVX512 support due to:
ModuleNotFoundError("No module named 'faiss.swigfaiss_avx512'")

Provider List: https://docs.litellm.ai/docs/providers
2025-02-18 06:31:35 INFO     faiss.loader  - Loading faiss with AVX2 support.

2025-02-18 06:31:35 INFO     faiss.loader  - Successfully loaded faiss with AVX2 support.
2025-02-18 06:31:35 INFO     faiss  - Failed to load GPU Faiss: name 'GpuIndexIVFFlat' is not defined. Will not load constructor refs for GPU indexes.