langroid 0.13.0__tar.gz → 0.15.0__tar.gz

{langroid-0.13.0 → langroid-0.15.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: langroid
-Version: 0.13.0
+Version: 0.15.0
 Summary: Harness LLMs with Multi-Agent Programming
 License: MIT
 Author: Prasad Chalasani
@@ -38,6 +38,7 @@ Provides-Extra: vecdbs
 Requires-Dist: aiohttp (>=3.9.1,<4.0.0)
 Requires-Dist: async-generator (>=1.10,<2.0)
 Requires-Dist: bs4 (>=0.0.1,<0.0.2)
+Requires-Dist: cerebras-cloud-sdk (>=1.1.0,<2.0.0)
 Requires-Dist: chainlit (==1.1.202) ; extra == "all" or extra == "chainlit"
 Requires-Dist: chromadb (>=0.4.21,<=0.4.23) ; extra == "vecdbs" or extra == "all" or extra == "chromadb"
 Requires-Dist: colorlog (>=6.7.0,<7.0.0)
@@ -153,6 +154,8 @@ This Multi-Agent paradigm is inspired by the
 `Langroid` is a fresh take on LLM app-development, where considerable thought has gone 
 into simplifying the developer experience; it does not use `Langchain`.
 
+:fire: Read the (WIP) [overview of the langroid architecture](https://langroid.github.io/langroid/blog/2024/08/15/overview-of-langroids-multi-agent-architecture-prelim/)
+
 📢 Companies are using/adapting Langroid in **production**. Here is a quote:
 
 >[Nullify](https://www.nullify.ai) uses AI Agents for secure software development. 

{langroid-0.13.0 → langroid-0.15.0}/README.md

@@ -46,6 +46,8 @@ This Multi-Agent paradigm is inspired by the
 `Langroid` is a fresh take on LLM app-development, where considerable thought has gone 
 into simplifying the developer experience; it does not use `Langchain`.
 
+:fire: Read the (WIP) [overview of the langroid architecture](https://langroid.github.io/langroid/blog/2024/08/15/overview-of-langroids-multi-agent-architecture-prelim/)
+
 📢 Companies are using/adapting Langroid in **production**. Here is a quote:
 
 >[Nullify](https://www.nullify.ai) uses AI Agents for secure software development. 

{langroid-0.13.0 → langroid-0.15.0}/langroid/agent/special/doc_chat_agent.py

@@ -14,6 +14,7 @@ pip install "langroid[hf-embeddings]"
 """
 
 import logging
+from collections import OrderedDict
 from functools import cache
 from typing import Any, Dict, List, Optional, Set, Tuple, no_type_check
 
@@ -130,12 +131,16 @@ class DocChatAgentConfig(ChatAgentConfig):
     n_fuzzy_neighbor_words: int = 100  # num neighbor words to retrieve for fuzzy match
     use_fuzzy_match: bool = True
     use_bm25_search: bool = True
+    use_reciprocal_rank_fusion: bool = True  # ignored if using cross-encoder reranking
     cross_encoder_reranking_model: str = (
         "cross-encoder/ms-marco-MiniLM-L-6-v2" if has_sentence_transformers else ""
     )
     rerank_diversity: bool = True  # rerank to maximize diversity?
     rerank_periphery: bool = True  # rerank to avoid Lost In the Middle effect?
     rerank_after_adding_context: bool = True  # rerank after adding context window?
+    # RRF (Reciprocal Rank Fusion) score = 1/(rank + reciprocal_rank_fusion_constant)
+    # see https://learn.microsoft.com/en-us/azure/search/hybrid-search-ranking#how-rrf-ranking-works
+    reciprocal_rank_fusion_constant: float = 60.0
     cache: bool = True  # cache results
     debug: bool = False
     stream: bool = True  # allow streaming where needed
@@ -1105,10 +1110,17 @@ class DocChatAgent(ChatAgent):
         Returns:
 
         """
-        # if we are using cross-encoder reranking, we can retrieve more docs
-        # during retrieval, and leave it to the cross-encoder re-ranking
-        # to whittle down to self.config.parsing.n_similar_docs
-        retrieval_multiple = 1 if self.config.cross_encoder_reranking_model == "" else 3
+        # if we are using cross-encoder reranking or reciprocal rank fusion (RRF),
+        # we can retrieve more docs during retrieval, and leave it to the cross-encoder
+        # or RRF reranking to whittle down to self.config.parsing.n_similar_docs
+        retrieval_multiple = (
+            1
+            if (
+                self.config.cross_encoder_reranking_model == ""
+                and not self.config.use_reciprocal_rank_fusion
+            )
+            else 3
+        )
 
         if self.vecdb is None:
             raise ValueError("VecDB not set")
@@ -1120,28 +1132,98 @@ class DocChatAgent(ChatAgent):
                     q,
                     k=self.config.parsing.n_similar_docs * retrieval_multiple,
                 )
+                # sort by score descending
+                docs_and_scores = sorted(
+                    docs_and_scores, key=lambda x: x[1], reverse=True
+                )
+
         # keep only docs with unique d.id()
-        id2doc_score = {d.id(): (d, s) for d, s in docs_and_scores}
-        docs_and_scores = list(id2doc_score.values())
-        passages = [d for (d, _) in docs_and_scores]
-        # passages = [
-        #     Document(content=d.content, metadata=d.metadata)
-        #     for (d, _) in docs_and_scores
-        # ]
+        id2_rank_semantic = {d.id(): i for i, (d, _) in enumerate(docs_and_scores)}
+        id2doc = {d.id(): d for d, _ in docs_and_scores}
+        # make sure we get unique docs
+        passages = [id2doc[id] for id, _ in id2_rank_semantic.items()]
 
+        id2_rank_bm25 = {}
         if self.config.use_bm25_search:
             # TODO: Add score threshold in config
             docs_scores = self.get_similar_chunks_bm25(query, retrieval_multiple)
-            passages += [d for (d, _) in docs_scores]
+            if self.config.cross_encoder_reranking_model == "":
+                # only if we're not re-ranking with a cross-encoder,
+                # we collect these ranks for Reciprocal Rank Fusion down below.
+                docs_scores = sorted(docs_scores, key=lambda x: x[1], reverse=True)
+                id2_rank_bm25 = {d.id(): i for i, (d, _) in enumerate(docs_scores)}
+                id2doc.update({d.id(): d for d, _ in docs_scores})
+            else:
+                passages += [d for (d, _) in docs_scores]
 
+        id2_rank_fuzzy = {}
         if self.config.use_fuzzy_match:
             # TODO: Add score threshold in config
             fuzzy_match_doc_scores = self.get_fuzzy_matches(query, retrieval_multiple)
-            passages += [d for (d, _) in fuzzy_match_doc_scores]
+            if self.config.cross_encoder_reranking_model == "":
+                # only if we're not re-ranking with a cross-encoder,
+                # we collect these ranks for Reciprocal Rank Fusion down below.
+                fuzzy_match_doc_scores = sorted(
+                    fuzzy_match_doc_scores, key=lambda x: x[1], reverse=True
+                )
+                id2_rank_fuzzy = {
+                    d.id(): i for i, (d, _) in enumerate(fuzzy_match_doc_scores)
+                }
+                id2doc.update({d.id(): d for d, _ in fuzzy_match_doc_scores})
+            else:
+                passages += [d for (d, _) in fuzzy_match_doc_scores]
 
-        # keep unique passages
-        id2passage = {p.id(): p for p in passages}
-        passages = list(id2passage.values())
+        if (
+            self.config.cross_encoder_reranking_model == ""
+            and self.config.use_reciprocal_rank_fusion
+            and (self.config.use_bm25_search or self.config.use_fuzzy_match)
+        ):
+            # Since we're not using cross-enocder re-ranking,
+            # we need to re-order the retrieved chunks from potentially three
+            # different retrieval methods (semantic, bm25, fuzzy), where the
+            # similarity scores are on different scales.
+            # We order the retrieved chunks using Reciprocal Rank Fusion (RRF) score.
+            # Combine the ranks from each id2doc_rank_* dict into a single dict,
+            # where the reciprocal rank score is the sum of
+            # 1/(rank + self.config.reciprocal_rank_fusion_constant).
+            # See https://learn.microsoft.com/en-us/azure/search/hybrid-search-ranking
+            #
+            # Note: diversity/periphery-reranking below may modify the final ranking.
+            id2_reciprocal_score = {}
+            for id_ in (
+                set(id2_rank_semantic.keys())
+                | set(id2_rank_bm25.keys())
+                | set(id2_rank_fuzzy.keys())
+            ):
+                rank_semantic = id2_rank_semantic.get(id_, float("inf"))
+                rank_bm25 = id2_rank_bm25.get(id_, float("inf"))
+                rank_fuzzy = id2_rank_fuzzy.get(id_, float("inf"))
+                c = self.config.reciprocal_rank_fusion_constant
+                reciprocal_fusion_score = (
+                    1 / (rank_semantic + c) + 1 / (rank_bm25 + c) + 1 / (rank_fuzzy + c)
+                )
+                id2_reciprocal_score[id_] = reciprocal_fusion_score
+
+            # sort the docs by the reciprocal score, in descending order
+            id2_reciprocal_score = OrderedDict(
+                sorted(
+                    id2_reciprocal_score.items(),
+                    key=lambda x: x[1],
+                    reverse=True,
+                )
+            )
+            # each method retrieved up to retrieval_multiple * n_similar_docs,
+            # so we need to take the top n_similar_docs from the combined list
+            passages = [
+                id2doc[id]
+                for i, (id, _) in enumerate(id2_reciprocal_score.items())
+                if i < self.config.parsing.n_similar_docs
+            ]
+            # passages must have distinct ids
+            assert len(passages) == len(set([d.id() for d in passages])), (
+                f"Duplicate passages in retrieved docs: {len(passages)} != "
+                f"{len(set([d.id() for d in passages]))}"
+            )
 
         if len(passages) == 0:
             return []
@@ -1171,7 +1253,7 @@ class DocChatAgent(ChatAgent):
             passages_scores = self.add_context_window(passages_scores)
             passages = [p for p, _ in passages_scores]
 
-        return passages
+        return passages[: self.config.parsing.n_similar_docs]
 
     @no_type_check
     def get_relevant_extracts(self, query: str) -> Tuple[str, List[Document]]:

{langroid-0.13.0 → langroid-0.15.0}/langroid/language_models/openai_gpt.py

@@ -21,6 +21,7 @@ from typing import (
 )
 
 import openai
+from cerebras.cloud.sdk import AsyncCerebras, Cerebras
 from groq import AsyncGroq, Groq
 from httpx import Timeout
 from openai import AsyncOpenAI, OpenAI
@@ -371,8 +372,8 @@ class OpenAIGPT(LanguageModel):
     Class for OpenAI LLMs
     """
 
-    client: OpenAI | Groq
-    async_client: AsyncOpenAI | AsyncGroq
+    client: OpenAI | Groq | Cerebras
+    async_client: AsyncOpenAI | AsyncGroq | AsyncCerebras
 
     def __init__(self, config: OpenAIGPTConfig = OpenAIGPTConfig()):
         """
@@ -479,6 +480,7 @@ class OpenAIGPT(LanguageModel):
             self.api_key = DUMMY_API_KEY
 
         self.is_groq = self.config.chat_model.startswith("groq/")
+        self.is_cerebras = self.config.chat_model.startswith("cerebras/")
 
         if self.is_groq:
             self.config.chat_model = self.config.chat_model.replace("groq/", "")
@@ -489,6 +491,16 @@ class OpenAIGPT(LanguageModel):
             self.async_client = AsyncGroq(
                 api_key=self.api_key,
             )
+        elif self.is_cerebras:
+            self.config.chat_model = self.config.chat_model.replace("cerebras/", "")
+            self.api_key = os.getenv("CEREBRAS_API_KEY", DUMMY_API_KEY)
+            self.client = Cerebras(
+                api_key=self.api_key,
+            )
+            # TODO there is not async client, so should we do anything here?
+            self.async_client = AsyncCerebras(
+                api_key=self.api_key,
+            )
         else:
             self.client = OpenAI(
                 api_key=self.api_key,
@@ -1096,8 +1108,8 @@ class OpenAIGPT(LanguageModel):
         if self.config.use_chat_for_completion:
             return self.chat(messages=prompt, max_tokens=max_tokens)
 
-        if self.is_groq:
-            raise ValueError("Groq does not support pure completions")
+        if self.is_groq or self.is_cerebras:
+            raise ValueError("Groq, Cerebras do not support pure completions")
 
         if settings.debug:
             print(f"[grey37]PROMPT: {escape(prompt)}[/grey37]")
@@ -1174,8 +1186,8 @@ class OpenAIGPT(LanguageModel):
         if self.config.use_chat_for_completion:
             return await self.achat(messages=prompt, max_tokens=max_tokens)
 
-        if self.is_groq:
-            raise ValueError("Groq does not support pure completions")
+        if self.is_groq or self.is_cerebras:
+            raise ValueError("Groq, Cerebras do not support pure completions")
 
         if settings.debug:
             print(f"[grey37]PROMPT: {escape(prompt)}[/grey37]")

{langroid-0.13.0 → langroid-0.15.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "langroid"
-version = "0.13.0"
+version = "0.15.0"
 description = "Harness LLMs with Multi-Agent Programming"
 authors = ["Prasad Chalasani <pchalasani@gmail.com>"]
 readme = "README.md"
@@ -89,6 +89,7 @@ async-generator = "^1.10"
 
 python-magic = "^0.4.27"
 json-repair = "^0.27.0"
+cerebras-cloud-sdk = "^1.1.0"
 
 
 [tool.poetry.extras]