Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
DETAILED ACTION
Claims 1-20 are pending. Claim 1 is independent.
This Application was published as US 20260187369.
Apparent priority is 30 December 2024.
The instant Application is directed to a method of providing a quality score with a RAG system.
Claim Objections
Claims 5, 9-10, 12-13, and 16-17 objected to because of the following informalities: in line 2 of each claim, “computer-implemented further comprises” is understood to mean “computer-implemented method further comprises”. Dependent claims 6-8, 11, 14-15 are also objected for the same reason as the claims that they depend upon. Appropriate correction is required.
Claim Rejections - 35 USC § 102
The following is a quotation of the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless –
(a)(2) the claimed invention was described in a patent issued under section 151, or in an application for patent published or deemed published under section 122(b), in which the patent or application, as the case may be, names another inventor and was effectively filed before the effective filing date of the claimed invention.
Claim(s) 1-6, 13, and 18-20 is/are rejected under 35 U.S.C. 102(a)(2) as being anticipated by Ortega de Castro Ayres et al. (US 20260147798 A1), hereinafter as Ortega.
Regarding claim 1, Ortega discloses: 1. A computer-implemented method for interacting with a retrieval-augmented generation system, comprising: ("[0001] Embodiments disclosed herein generally relate to detecting hallucinations in systems that include machine learning models including retrieval augmented generation (RAG) based systems…" )
receiving a textual prompt; ("[0047] As previously stated, FIG. 1 discloses aspects of the hallucination detection system 100 deployed in a LLM-based application using RAG. In this example, the LLM-based application is a question/answer application. FIG. 1 assumes that a query has been submitted to the application and that an answer has been generated and/or that sources for the answer are identified." )
retrieving information associated with the textual prompt based at least in part on the textual prompt; ("[0047] As previously stated, FIG. 1 discloses aspects of the hallucination detection system 100 deployed in a LLM-based application using RAG. In this example, the LLM-based application is a question/answer application. FIG. 1 assumes that a query has been submitted to the application and that an answer has been generated and/or that sources for the answer are identified." )
generating input data based at least in part on the textual prompt and the retrieved information; ("[0046]... The context 204 identifies sources from that are used to provide context to a prompt. The sources may be identified by a RAG system..." )
generating an output at least in part by applying the input data to a machine learning model of the retrieval-augmented generation system, ("[0047] As previously stated, FIG. 1 discloses aspects of the hallucination detection system 100 deployed in a LLM-based application using RAG. In this example, the LLM-based application is a question/answer application. FIG. 1 assumes that a query has been submitted to the application and that an answer has been generated and/or that sources for the answer are identified.")
the machine learning model configured using at least prompt engineering to determine the output based at least in part on the input data; ("[0004] This technique may also be adopted to prevent hallucinations. More specifically, the LLM can receive instructions to answer questions or queries using only the contextual information of a list of sources that are placed within the prompt…" )
outputting the output and ("[0050] In FIG. 1, an output, the answer 122, of the RAG answer builder 102 is received at the hallucination detection system 100 and may be formatted as a tuple…" – see also “[0048] Thus, the hallucination detection module 100 is positioned, in one example, to determine whether the answer is a hallucination prior to providing the answer to the user…”)
an indication of the quality score. ("[0043] In one example, a set of features are built to train a hallucination detection model. These features may combine natural language processing (NLP) techniques with specifically designed features to address the unique aspects of hallucinations. This feature set is employed to develop/train a machine learning model that outputs hallucination scores. Users may adjust and set hallucination thresholds for their specific circumstances." )
Regarding claim 2, Ortega discloses: 2. The computer-implemented method of claim 1, wherein the indication of the quality score comprises: the quality score; ("[0043] In one example, a set of features are built to train a hallucination detection model. These features may combine natural language processing (NLP) techniques with specifically designed features to address the unique aspects of hallucinations. This feature set is employed to develop/train a machine learning model that outputs hallucination scores. Users may adjust and set hallucination thresholds for their specific circumstances." - see also: "[0080] As illustrated in FIG. 1, the hallucination detection system 100 is integrated in an LLM-based system using RAG. The system 100 is called, in one example, after the answer builder 102 and returns a score or result between 0 and 100%. Higher scores indicate a higher likelihood that the answer is a hallucination. The score may depend on the threshold or level of hallucination set by a user or set in another manner." )
a rating derived from the quality score; an indicator with a colour corresponding to the quality score; and/or an indicator with a colour corresponding to the rating derived from the quality score.
("[0045] The hallucination dataset 130 may be an existing dataset, an internal dataset, or the like. In one example, the dataset 130 may include elements or entries formatted as tuples. The format of the tuples is (query, list of sources to be used as context, answer). A hallucination flag that can be true when the answer is considered to be a hallucination or false when the answer is covered by the content of the list of sources, and is not a hallucination. Alternatively, examples of hallucination tuples with hallucination flag equal to true can be artificially generated by shuffling answers of domain specific datasets that contain a query, a list of prompt sources and its respective LLM answer." – a hallucination flag reads on a rating. As cited in [0080] above the user can set a threshold for the hallucination score.)
Regarding claim 3, Ortega discloses: 3. The computer-implemented method of claim 2, wherein: retrieving information associated with the textual prompt comprises retrieving an electronic file comprising text associated with the textual prompt; and the output comprises a textual output. (Fig. 2 shows examples that the context and answer are textual. )
Regarding claim 4, Ortega discloses: 4. The computer-implemented method of claim 3, wherein the quality score is associated with: a lexical similarity between the textual output and related text of the retrieved information; and/or a semantic similarity between the textual output and related text of the retrieved information. ("[0087] Embodiments of the invention relate to a hallucination detection system that uses novel features combining lexical, semantic and predefined rules to detect hallucinations in LLMs that use RAG. In one example, the best features are selected based on, by way of example, the machine learning classifier, the objective function based on common metrics, and applied to a specific hallucination dataset." )
Regarding claim 5, Ortega discloses: 5. The computer-implemented method of claim 4, wherein: the computer-implemented further comprises generating a lexical score associated with the lexical similarity; and ("[0022] Various methods may be used to determine the similarity of texts or between pairs of texts. Example methods include lexical methods and semantic methods. Lexical methods relate to the word frequency or word overlapping between pairs of texts. Semantic methods capture the semantic meaning of sentences, which is distinct from considering only the raw words as in lexical methods." )
the quality score is generated based at least in part on the lexical score. ([0087] discloses that lexical features are used to generate the score. )
Regarding claim 6, Ortega discloses: 6. The computer-implemented method of claim 5, wherein the lexical score is generated based at least in part on calculating Jaccard similarity or F1 score. ("[0076] The configuration possibilities include the list of potential features explored in the feature engineering, the machine learning classification algorithms (e.g., Decision Trees, Gaussian Process, Random Forest, SVC, Naïve Bayes, a list of machine learning metrics such as precision, recall, F1-score, accuracy, and the hallucination data set selected." )
Regarding claim 13, Ortega discloses: 13. The computer-implemented method of claim 4, wherein: the computer-implemented further comprises generating a lexical score associated with the lexical similarity and generating a semantic score associated with the semantic similarity; and the quality score is generated based at least in part on the lexical score and the semantic score. ("[0087] Embodiments of the invention relate to a hallucination detection system that uses novel features combining lexical, semantic and predefined rules to detect hallucinations in LLMs that use RAG. In one example, the best features are selected based on, by way of example, the machine learning classifier, the objective function based on common metrics, and applied to a specific hallucination dataset." )
Regarding claim 18, Ortega discloses: 18. The computer-implemented method of claim 1, wherein the machine learning model comprises a language model. ("[0014] Embodiments disclosed herein generally relate to detecting hallucinations in machine learning models and systems. More particularly, at least some embodiments relate to systems, hardware, software, computer-readable media, and methods for detecting hallucinations in large language model (LLM) based applications that include retrieval augmented generation (RAG)." )
Regarding claim 19, Ortega discloses: 19. The computer-implemented method of claim 1, wherein the machine learning model comprises a generative language model. ("[0014] Embodiments disclosed herein generally relate to detecting hallucinations in machine learning models and systems. More particularly, at least some embodiments relate to systems, hardware, software, computer-readable media, and methods for detecting hallucinations in large language model (LLM) based applications that include retrieval augmented generation (RAG)." )
Regarding claim 20, Ortega discloses: 20. A system comprising: one or more processors; and memory storing a computer program configured to be executed by the one or more processors; wherein the computer program comprises instructions for performing or facilitating performing of the computer-implemented method of claim 1. ("[0111] The embodiments disclosed herein may include the use of a special purpose or general-purpose computer including various computer hardware or software modules, as discussed in greater detail below. A computer may include a processor and computer storage media carrying instructions that, when executed by the processor and/or caused to be executed by the processor, perform any one or more of the methods disclosed herein, or any part(s) of any method disclosed." )
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claim(s) 7-12 is/are rejected under 35 U.S.C. 103 as being unpatentable over Ortega in view of Banzhaf et al. (US 12645686 B1).
Regarding claim 7, Ortega discloses: 7. The computer-implemented method of claim 6, wherein generating the lexical score comprises: filtering stop words from the textual output and the related text of the retrieved information; (Not explicitly disclosed )
lemmatizing the filtered output and the filtered related text of the retrieved information; and (Not explicitly disclosed )
generating the lexical score at least in part by calculating an F1 score associated with the lemmatized and filtered output and the lemmatized and filtered related text of the retrieved information. ("[0032] The ROUGE-1 F1-score can be directly obtained from the ROUGE-1 precision and the ROUGE-1 recall using the standard F1-score formula:" )
Ortega does not explicitly disclose stop word filtering or lemmatization.
Banzhaf discloses: filtering stop words from the textual output and the related text of the retrieved information; lemmatizing the filtered output and the filtered related text of the retrieved information; ("For example, the machine learning model may tokenize the query, perform processing such as stop word removal, stemming, and/or lemmatization to reduce words to their base form, and remove common words that do not add to the query..." Col 14, para 2)
Ortega and Banzhaf are considered analogous art to the claimed invention because they disclose systems that interact with RAG. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the system of Ortega with stop word filtering and lemmatization as disclosed by Banzhaf. Doing so would have been beneficial to remove common words that do not add to the query and to reduce words to their base form. (Banzhaf Col 14, para 2). This combination falls under combining prior art elements according to known methods to yield predictable results or use of known technique to improve similar devices (methods, or products) in the same way. See MPEP 2141, KSR, 550 U.S. at 418, 82 USPQ2d at 1396.
Regarding claim 8, Ortega discloses: 8. The computer-implemented method of claim 7, wherein the calculating of the F1 score is based on:
(
2
×
p
r
e
c
i
s
i
o
n
×
r
e
c
a
l
l
)
p
r
e
c
i
s
i
o
n
+
r
e
c
a
l
l
("[0032] The ROUGE-1 F1-score can be directly obtained from the ROUGE-1 precision and the ROUGE-1 recall using the standard F1-score formula:
PNG
media_image1.png
53
348
media_image1.png
Greyscale
" )
where precision = W/R, recall = W/S, W corresponds to a number of overlapping words between the lemmatized and filtered output and the lemmatized and filtered related text of the retrieved information, R corresponds to a number of words in the lemmatized and filtered output, and S corresponds to a number of words in the lemmatized and filtered text of the retrieved information.
("[0029] An example of ROUGE metrics considers the following reference R and candidate summary C: R: The baby is in the crib. C: The baby and the mom. ROUGE-N [0030] ROUGE-N is an overlap of n-grams between the generated text and the reference. [0031] ROUGE-1 considers unigrams. ROUGE-1 precision can be computed as a ratio of the number of unigrams in C that also appear in R (the words “the”, “baby” and “the”), over the number of unigrams in C. Rouge-1 recall, on the other hand, compares the number of unigrams in C that appear also in R over the number of unigrams in R:…" )
Regarding claim 9, Ortega discloses: 9. The computer-implemented method of claim 8, wherein: the computer-implemented further comprises outputting an indication of the lexical score; and (Fig. 1 shows that the features calculated at 116 are output to the Classification model. )
the indication of the lexical score comprises: the lexical score; a rating derived from the lexical score; an indicator with a colour corresponding to the lexical score; and/or an indicator with a colour corresponding to the rating derived from the lexical score. ("[0059] The system 100 calculates 116 features for the answer 124. The features that are calculated may be a result of feature engineering. The classification model at 118 may also be trained using the features calculated at 116." - using the score as a feature reads on the indication being the lexical score. )
Regarding claim 10, Ortega discloses: 10. The computer-implemented method of claim 9, wherein: the computer-implemented further comprises generating a semantic score associated with the semantic similarity; and ("[0022] Various methods may be used to determine the similarity of texts or between pairs of texts. Example methods include lexical methods and semantic methods. Lexical methods relate to the word frequency or word overlapping between pairs of texts. Semantic methods capture the semantic meaning of sentences, which is distinct from considering only the raw words as in lexical methods." )
the quality score is generated based at least in part on the semantic score. ([0087] discloses that semantic features are used to generate the score.)
Regarding claim 11, Ortega discloses: 11. The computer-implemented method of claim 10, wherein generating the semantic score comprises: generating the semantic score at least in part by calculating a cosine similarity between embeddings of the textual output and embeddings of the related text of the retrieved information. ("[0040] Semantic similarity methods can be calculated using the similarity of their embeddings. The embedding is obtained at the output of the encoder block of a transformer architecture. An encoder, give an input text, generates an output that includes a high dimensional array that captures the semantic information of the text. Semantically comparing two texts includes generating their embeddings and then measuring their similarity mathematically (e.g., cosine similarity, Euclidean distance or dot product)." )
Regarding claim 12, Ortega discloses: 12. The computer-implemented method of claim 11, wherein: the computer-implemented further comprises outputting an indication of the semantic score; and (Fig. 1 shows that the features calculated at 116 are output to the Classification model.)
the indication of the semantic score comprises: the semantic score; a rating derived from the semantic score; an indicator with a colour corresponding to the semantic score; and/or an indicator with a colour corresponding to the rating derived from the semantic score. ("[0059] The system 100 calculates 116 features for the answer 124. The features that are calculated may be a result of feature engineering. The classification model at 118 may also be trained using the features calculated at 116." - using the score as a feature reads on the indication being the score itself. )
Claim(s) 14 is/are rejected under 35 U.S.C. 103 as being unpatentable over Ortega in view of Mukherjee et al. (US 20260093734 A1).
Regarding claim 14, Ortega does not disclose the additional limitations.
Mukherjee discloses: 14. The computer-implemented method of claim 13, wherein the quality score is generated based at least in part on: w1 ‧ (lexical score) + w2 ‧ (semantic score) where w1 is a weighting for the lexical score and w2 is a weighting for the semantic score. ("[0055] In some examples, a hybrid search of lexical searching and semantic searching is supported. In some examples, lexical retrieval scores using min-max normalization and a weighted average of the lexical and semantic scores is calculated to determine a final score." )
Ortega and Mukherjee are considered analogous art to the claimed invention because they disclose systems that interact with RAG systems. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the system of Ortega with a weighted average lexical and semantic score as disclosed by Mukherjee. Doing so would have been beneficial so both lexical and semantic scores are taken into account. (Mukherjee, [0055]). This combination falls under combining prior art elements according to known methods to yield predictable results or use of known technique to improve similar devices (methods, or products) in the same way. See MPEP 2141, KSR, 550 U.S. at 418, 82 USPQ2d at 1396.
Claim(s) 15 is/are rejected under 35 U.S.C. 103 as being unpatentable over Ortega in view of Mukherjee as applied in claim 14 above, further in view of MathIsFun ("Weighted Mean").
Regarding claim 15, Ortega and Mukherjee do not disclose the additional limitations.
MathIsFun discloses: 15. The computer-implemented method of claim 14, wherein:
quality score = w1 ‧ (lexical score) + w2 ‧ (semantic score) and w1 + w2 = 1. ("When the weights add to 1: just multiply each weight by the matching value and sum it all up" pg. 6, first line)
Ortega and Mukherjee are considered analogous art to the claimed invention because they disclose systems that interact with RAG systems. MathIsFun is considered analogous art to the claimed invention because they disclose methods of calculating weighted averages. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the combination with a weighted average where the weights sum to one as disclosed by MathIsFun. Doing so would have been beneficial to avoid an extra division step to get an average. (MathIsFun, pg. 6, second bullet). This combination falls under combining prior art elements according to known methods to yield predictable results or use of known technique to improve similar devices (methods, or products) in the same way. See MPEP 2141, KSR, 550 U.S. at 418, 82 USPQ2d at 1396.
Claim(s) 17 is/are rejected under 35 U.S.C. 103 as being unpatentable over Ortega in view of Tiwari et al. (US 20260093931 A1).
Regarding claim 17, Ortega discloses: 17. The computer-implemented method of claim 1, wherein: the computer-implemented further comprises providing a user interface associated with the retrieval-augmented generation system; (Fig. 5 shows UI Device 510)
the textual prompt is received via the user interface; (“[0066]...Calculating this adherence scores indicates whether the context generated by the source manager module of the LLM-based system is aligned with the question the user entered...” - One of ordinary skill in the art would have understood that the user enters the query in the user interface.)
and the computer-implemented further comprises displaying the output (“[0048] Thus, the hallucination detection module 100 is positioned, in one example, to determine whether the answer is a hallucination prior to providing the answer to the user…”)
and the indication of the quality score in the user interface. (not explicitly disclosed)
Ortega does not explicitly disclose displaying the indication of the quality score in the user interface.
Tiwari discloses: displaying the indication of the quality score in the user interface. ("[0188] At block 608, the one or more processors provide an indication of the confidence score for display on a user interface. The confidence score indicates the reliability of the response... " )
Ortega and Tiwari are considered analogous art to the claimed invention because they disclose systems that interact with RAG systems. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the system of Ortega with displaying the indication of the quality score in the user interface. Doing so would have been beneficial to indicate the reliability of the response. (Tiwari [0005]).
Claim(s) 16 is/are rejected under 35 U.S.C. 103 as being unpatentable over Ortega in view of Mukherjee and MathIsFun as applied in claim 15 above, further in view of Tiwari.
Regarding claim 16, Ortega discloses: 16. The computer-implemented method of claim 15, wherein: the computer-implemented further comprises providing a user interface associated with the retrieval-augmented generation system; (Fig. 5 shows UI Device 510 )
the textual prompt is received via the user interface; and ("[0066]...Calculating this adherence scores indicates whether the context generated by the source manager module of the LLM-based system is aligned with the question the user entered..." - One of ordinary skill in the art would have understood that the user enters the query in the user interface.)
the computer-implemented further comprises outputting an indication of the lexical score and an indication of the semantic score, and (Fig. 1 shows that the features calculated at 116 are output to the Classification model. )
displaying the output, (“[0048] Thus, the hallucination detection module 100 is positioned, in one example, to determine whether the answer is a hallucination prior to providing the answer to the user…”)
the indication of the quality score, the indication of the lexical score, and the indication of the semantic score in the user interface. (Not explicitly disclosed )
Ortega does not explicitly disclose displaying the indications of the scores in the user interface. Neither does Mukherjee or MathIsFun.
Tiwari discloses: displaying the indication of the scores in the user interface. ("[0188] At block 608, the one or more processors provide an indication of the confidence score for display on a user interface. The confidence score indicates the reliability of the response... " )
Ortega, Mukherjee, and Tiwari are considered analogous art to the claimed invention because they disclose systems that interact with RAG systems. MathIsFun is considered analogous art to the claimed invention because they disclose methods of calculating weighted averages. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the system of Ortega with displaying the indications of the quality scores in the user interface. Doing so would have been beneficial to indicate the reliability of the response. (Tiwari [0005]).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Erickson (US 20240386041 A1). Erickson discloses a method of private RAG with prompt engineering. (Fig. 1B)
Any inquiry concerning this communication or earlier communications from the examiner should be directed to JON C MEIS whose telephone number is (703)756-1566. The examiner can normally be reached Monday - Thursday, 8:30 am - 5:30 pm EST.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Hai Phan can be reached at 571-272-6338. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/JON CHRISTOPHER MEIS/Examiner, Art Unit 2654
/HAI PHAN/Supervisory Patent Examiner, Art Unit 2654
Prosecution Insights