Tu'ufa'atasia o le RAG ma LLMs mo le Fa'ata'ita'iga Fa'atonu Fa'atonu: Ose Ta'iala Fa'atino | B4N1

Tu'ufa'atasia o le RAG ma LLMs mo le Fa'ata'ita'iga Fa'atonu Fa'atonu: Ose Ta'iala Fa'atino

8 months ago · b4n1 Read time: 5 min

** Tu'ufa'atasia le RAG ma LLMs mo le Fa'atulagaina o Polokalama Fa'ata'ita'i: O se Ta'iala Fa'atino**

Fa'atomuaga

I tausaga talu ai nei, o le fanua o le faʻapipiʻiina o masini komepiuta ua faia se suiga tele i le oʻo mai o Large Language Models (LLMs) ma Retrieval-Augmented Generation (RAG). O le feso'ota'iga i le va o nei tekonolosi e lua e iai le avanoa e fa'afouina ai le auala tatou te mamanuina ma atina'e ai polokalama faakomepiuta. I totonu o lenei tusiga, o le a matou suʻesuʻeina le tuʻufaʻatasia o le RAG ma LLMs mo le faʻaogaina o masini komepiuta, e aofia ai taʻiala faʻatinoina e faʻaaoga ai meafaigaluega faʻapipiʻi e pei o LangChain ma DSPy.

O le a le RAG?

O le RAG o se metotia e tu'ufa'atasia ai le malosi o le Retrieval-based ma Generation-based fa'ata'ita'iga e maua mai ai tusitusiga maualuga. O le manatu autu i tua o le RAG o le faʻaaogaina lea o se faʻataʻitaʻiga toe maua e filifili ai pepa talafeagai poʻo fuaitau mai se tino tele, ona faʻaaoga lea o se faʻataʻitaʻiga faʻatuputupulaga e faʻatupu ai tusitusiga e faʻavae i luga o nei fuaitau na maua mai. O lenei faiga ua faʻaalia ai le faʻaleleia atili o galuega e pei o le taliina o fesili, otootoga o tusitusiga, ma talanoaga AI.

O le a le LLM?

LLMs o se ituaiga o fa'ata'ita'iga fa'ata'ita'i fa'ata'ita'i e a'oa'oina i le tele o fa'amaumauga o tusitusiga e fa'atupu ai fa'atinoga o gagana e pei o tagata. O nei faʻataʻitaʻiga e masani lava ona faʻavae i luga o faʻataʻitaʻiga o le transformer ma ua ausia taunuʻuga faʻaonaponei i le tele o galuega faʻaogaina gagana masani. LLMs e sili ona fa'aogaina ma e mafai ona fa'aoga mo galuega e pei o le fa'aliliuga gagana, fa'apotopotoga o tusitusiga, ma talanoaga AI.

Aisea e Tu'ufa'atasi ai le RAG ma LLMs?

O le tuʻufaʻatasia o le RAG ma LLM e ofoina atu le tele o faʻamanuiaga mo le faʻaogaina o masini komepiuta:

  1. Fa'aleleia le Sa'o: E mafai e le RAG ona fesoasoani i LLM e filifili pepa talafeagai po'o fuaitau mai se tino tele, e o'o atu ai i fa'amatalaga sa'o ma fa'amatalaga.
  2. Fa'ateleina le Lelei: E mafai e le RAG ona fa'aitiitia le fa'asologa o fa'asologa o LLM e ala i le filifilia na'o pepa e sili ona talafeagai po'o fuaitau, e o'o atu ai i taimi vave tali.
  3. Enhanced Flexibility: E mafai ona fa'aoga le RAG e fa'afetaui ai le suiga o mana'oga ma mana'oga o tagata, fa'atagaina LLM e fa'atupu fa'amatalaga tusitusia e sili atu ona talafeagai ma fa'aoga.

Ta'iala Fa'atino

E tele ta'iala fa'atino mo le tu'ufa'atasia o le RAG ma LLM, e aofia ai:

  1. LangChain: LangChain o se faletusi matala e maua ai se API faigofie ma faʻaogaina mo le fausiaina o faʻataʻitaʻiga faʻavae RAG. E lagolagoina le tele o LLM ma e mafai ona faʻaaogaina e faʻatino ai faʻataʻitaʻiga RAG lavelave.
  2. ** DSPy **: DSPy o se faletusi Python e maua ai se auala faigofie ma lelei e faʻatino ai faʻataʻitaʻiga faʻavae RAG. E lagolagoina le tele o LLM ma e mafai ona faʻaaogaina e faʻatino ai faʻataʻitaʻiga RAG lavelave.

Fa'ata'ita'iga Fa'ailoga

O se faʻataʻitaʻiga faigofie lea e faʻaalia ai le faʻaogaina o le LangChain e tuʻufaʻatasia ai le RAG ma LLM:

faaulufale mai le langchain
mai langchain.chains faaulufale mai RAGCain
mai langchain.llms faaulufale mai LLaMA


llama = LLaMA()

# Fausia se faʻataʻitaʻiga RAGCain
rag_chain = RAGCain(llama, retrieval_model="t5-small")

# Faʻamatala se faʻamatalaga
vave = "O le a le laumua o Farani?"

# Fausia mea e maua ai tusitusiga
gaosiga = rag_chain.run(vave)

lolomi(output)

O lenei faʻataʻitaʻiga faʻataʻitaʻiga o loʻo faʻaalia ai le faʻaogaina o LangChain e fatu ai se faʻataʻitaʻiga RAGChain e tuʻufaʻatasia se faʻataʻitaʻiga LLaMA ma se faʻataʻitaʻiga toe maua. O le faʻataʻitaʻiga RAGCain e faʻaaogaina e gaosia ai tusitusiga e faʻavae i luga o se vave.

Poloaiga/Snippets pito i luga 5-10 Fa'aaogaina

O le pito i luga 5-10 pito sili ona faʻaaogaina tulafono / snippets mo le tuʻufaʻatasia o le RAG ma LLMs:

  1. langchain.RAGCain(llama, retrieval_model="t5-small"): Fausia se faʻataʻitaʻiga RAGCain e tuʻufaʻatasia se faʻataʻitaʻiga LLaMA ma se faʻataʻitaʻiga toe maua.
  2. llama.run(prompt): Fausia tusitusiga e fa'atatau i se vave fa'aaoga le fa'ata'ita'iga LLaMA.
  3. rag_chain.run(prompt): Fausia tusitusiga e fa'atatau i se vave fa'aaoga le RAGChain fa'ata'ita'iga.
  4. langchain.chains.RAGCain(llama, retrieval_model="t5-small", max_length=1024): Fausia se faʻataʻitaʻiga RAGCain e tuʻufaʻatasia se faʻataʻitaʻiga LLaMA ma se faʻataʻitaʻiga toe maua ma faʻatulagaina le umi maualuga o le tusitusiga o loʻo gaosia.
  5. llama.get_input(prompt): Fa'afo'i le tusitusiga o lo'o i totonu mo le fa'ata'ita'iga LLaMA.
  6. rag_chain.get_output(prompt): Toe fa'afo'i le tusitusiga mo le ata RAGChain.
  7. langchain.chains.RAGChain(llama, retrieval_model="t5-small", batch_size=32): Fausia se faʻataʻitaʻiga RAGChain e tuʻufaʻatasia se faʻataʻitaʻiga LLaMA ma se faʻataʻitaʻiga toe faʻatulagaina ma seti le tele o le vaega.
  8. llama.get_output(prompt): Fa'afo'i le tusitusiga o lo'o gaosia mo le fa'ata'ita'iga LLaMA.
  9. rag_chain.get_input(prompt): Toe fa'afo'i le tusitusiga fa'aoga mo le fa'ata'ita'iga RAGChain.
  10. langchain.chains.RAGCain(llama, retrieval_model="t5-small", device="cuda"): Fausia se faʻataʻitaʻiga RAGCain e tuʻufaʻatasia se faʻataʻitaʻiga LLaMA ma se faʻataʻitaʻiga toe maua ma seti le masini i le CUDA.

Va'aiga

Ole fa'aaliga faigofie lea ole fa'ata'ita'iga RAGCain fa'aaoga Chart.js:

{
  "type": "pa",
  "fa'amatalaga": {
    "labels": ["LLaMA", "RAGChain"],
    "fa'amaumauga": [
      {
        "label": "Sa'o",
        "fa'amatalaga": [0.8, 0.9],
        "Lulu lanu": ["#FF6384", "#36A2EB"]
      }
    ]
  },
  "filifiliga": {
    "igoa": {
      "fa'aaliga": moni,
      "text": "RAGCain Fa'ata'ita'iga Sa'o"
    },
    "fua": {
      "yAxes": [
        {
          "tiki": {
            "beginAtZero": moni
          }
        }
      ]
    }
  }
}

O lenei faʻaaliga o loʻo faʻaalia ai le saʻo o le faʻataʻitaʻiga LLaMA ma le faʻataʻitaʻiga RAGChain.

Fa'ai'uga

o le tu'ufa'atasiga o le RAG ma LLM e ofoina atu ai le tele o fa'amanuiaga mo le fa'aogaina o masini komepiuta, e aofia ai le fa'aleleia atili o le sa'o, fa'ateleina le lelei, ma le fa'aleleia atili o le fetuutuuna'i. LangChain ma DSPy o ni meafaigaluega faʻapipiʻi taʻutaʻua e lua e mafai ona faʻaaogaina e faʻatino ai faʻataʻitaʻiga faʻavae RAG. E ala i le mulimuli i taʻiala faʻatinoga ma faʻataʻitaʻiga faʻataʻitaʻiga o loʻo tuʻuina atu i lenei tusiga, e mafai e tagata atiaʻe ona fatuina faʻataʻitaʻiga RAG lavelave e tuʻufaʻatasia LLM ma faʻataʻitaʻiga toe maua. Ole pito i luga ole 5-10 pito sili ona faʻaaogaina tulafono / snippets mo le tuʻufaʻatasia o le RAG ma LLM o loʻo tuʻuina atu foi mo faʻamatalaga.

Have Questions?

Let's Talk

Contact me on LinkedIn