Update rag-application-communication-system.md

small change

docs/articles/rag-application-communication-system.md

@@ -20,15 +20,15 @@ But LLMs are also limited. They are few doc-learners, largely conceived as “in
 * *Updatability:* while LLMs are limited by their cutoff date, *RAG systems can be continuously maintained in real time*.
 * *Explainability:* LLMs are a “black box”; *RAG systems can provide transparency, tracing statements to original sources*.
 
-The default paradigm and basic conceptualization of RAG emerged almost immediately after the release of chatGPT. Simon Willison’s “[semantic search answers pattern](https://simonwillison.net/2023/Jan/13/semantic-search-answers/)” (January 2023) is one of the first systematic descriptions: search for texts in a semantic database, get the most relevant results, send this input to GPT-3 as inspiration. Following this basic strategy, RAG helps mitigate the limitations of few document learning by grounding an LLM on existing data at inference time.
+The default paradigm and basic conceptualization of RAG emerged almost immediately after the release of ChatGPT. Simon Willison’s “[semantic search answers pattern](https://simonwillison.net/2023/Jan/13/semantic-search-answers/)” (January 2023) is one of the first systematic descriptions: search for texts in a semantic database, get the most relevant results, send this input to GPT-3 as inspiration. Following this basic strategy, RAG helps mitigate the limitations of few document learning by grounding an LLM on existing data at inference time.
 
 **But RAG doesn't solve LLM limitations once and for all**. *In production*:
 
 * a) there's *no one-size-fits-all RAG solution*, as people working on production RAG today soon come to realize. Good applications are increasingly a lot of flows and arrows, anticipating a wide array of edge cases. There are simply too many breaking points - e.g., bad OCR, poor indexation, suboptimal tokenization, inconsistent document segmentation, unpredictable queries, etc. - to settle on one generic approach.
 * b) *your solution to the last mile problem (indexing, retrieval, and generation) depends heavily on your use case*: This won't be easily handled by LLMs, despite major AI actors' eagerness to create innovative panacea platforms. Probably the only actual moat in finding the right fit for your use case, corpus, and specific user target is in applied generative AI - i.e., in a specifically and dynamically tailored RAG system.
-* c) *RAG evaluation is tricky*: While you can use synthetic data to do preliminary evaluations and benchmarks for selecting and swapping models, the real challenge is evaluation in production of your system's different components, taking into account their level of specialization. Though we'd like to automate everything, including evaluation, here it continues to require qualitative human involvement.
+* c) *RAG evaluation is tricky*: While you can use synthetic data to do preliminary evaluations and benchmarks for selecting and swapping models, the real challenge is evaluation in production of your system's different components, taking into account their level of specialization. Though we'd like to automate everything, including evaluation, here we continue to require qualitative human involvement.
 
-These ongoing difficulties are at once challenges *and* opportunities for RAG applications *qua* communication systems - employing LLMs in a particular landscape of specific data and users. In a way, this has always been RAG's *raison d'etre*. As communication systems, RAG applications can perform better than even frontier LLMs alone, by taking advantage of a broad range of techniques, not all of them new, as we'll see below.
+These ongoing difficulties are at the same time challenges *and* opportunities for RAG applications *qua* communication systems - employing LLMs in a particular landscape of specific data and users. In a way, this has always been RAG's *raison d'etre*. As communication systems, RAG applications can perform better than even frontier LLMs alone, by taking advantage of a broad range of techniques, not all of them new, as we'll see below.
 
 ## 2. Retrieval: many solutions - not everything new is better