Improvement of the AI assistant for the contact center operator of the medical center — ChatGPT, RAG, Skills

Budget: 27000 UAH Deadline: 12 days

Hello.

The problem of losing context in large Markdown files and missing medical combinations is solved by changing the chunking strategy and strict routing of sources. I will rewrite Python scripts to parse your CSVs into compact indexes, optimize SKILL.md, and set up precise RAG search without hallucinations regarding prices. I will ensure absolute accuracy in extracting specific data by applying architectural approaches from my experience in developing autonomous AI agents (OpenClaw) and B2B lead generation systems. What is the current token volume of your typical system prompt along with the base in Project Sources?

I am ready to sign an NDA and start the audit.

Budget: 27000 UAH Deadline: 30 days

I work with LLM assistants, SKILL.md, structured knowledge bases, Markdown/CSV/JSONL, and Python data generators.

I will start with an audit of the retrieval chain: file structure, block sizes, indexes, source routing, and mandatory search rules. Next, I will separate critical entities, set up the search for packages, combinations, prices, performers, and surgical cases, and create a set of control queries with quality criteria. I am ready to work under an NDA. Budget and deadlines will be discussed in private correspondence.

Do you allow the critical retrieval to be moved to an external index or API if testing shows the stability limit of Project Sources?

Budget: 5000 UAH Deadline: 7 days

Hello! I do exactly these things — generating a structured database from CSV and setting up search functionality for it. From the description, the likely reason for the gaps is clear: when the service itself, prices by subdivisions, and recommended combinations are located in different places within a large file, retrieval extracts the first match and stops — hence the base price of the operation is found, but the total cost of the case is not. This is resolved not by a prompt, but by restructuring the record: one service code = one self-contained block, where the price, subdivisions, executors, and related packages are already stitched together, plus a control set of queries to see regressions. I would start with an audit of SKILL.md and the generator from CSV. What is the current volume of files in Project Sources?

Budget: 5000 UAH Deadline: 5 days

Hello, Yevhen!

I read the brief carefully. The problem is rather not in the prompt itself, but in the retrieval from large Markdown sources: when the service and related data (combinations, cost of the surgical case, subdivisions) are located in different places within large files, the model retrieves the first relevant block and "does not read" the rest. This is a classic context fragmentation issue in ChatGPT Projects.

My stack is exactly this: I build LLM assistants on RAG and structured Skills (ChatGPT Skills / Custom GPT + my own compact indexes + Python knowledge base generators from CSV/JSON). I have done searches in catalogs with thousands of entries, ensuring the extraction of related fields.

How I diagnose missing fields: (1) I collect a set of control queries (like your "cholecystectomy") with benchmark answers; (2) I run them and see WHAT exactly the model is missing — this shows whether the problem lies in chunking files, the structure of indexes, or the routing of sources; (3) I rebuild the database on the principle "one service → all related fields in one record," so that retrieval does not have to stitch data from different files.

What I will do: audit SKILL.md and Project Sources → rewritten SKILL.md with mandatory search steps (packages, combinations, prices by subdivisions, performers, cost of the case) → optimized Markdown/JSONL indexes → refined Python generator from CSV → control queries + evaluation criteria → testing on real scenarios → final documentation.

Timeline: audit + first iteration of refinement ≈ 4-5 days. Cost — within your budget, we will finalize after clarifying the terms of reference, as you mentioned. NDA — ready to sign.

One clarification: is the database in Project Sources currently one large Markdown for everything, or is it already split into files (prices/performers/packages separately)? This affects whether to optimize the existing structure or rebuild the indexes.

Petro Pankov, BotCraft Group Expand

Budget: 5000 UAH Deadline: 7 days

Hello, I have worked on RAG assistants with large service catalogs and custom GPT on ChatGPT Projects. I built a search across 2000+ items from structured Markdown and JSONL, configured source routing and retrieval rules. I wrote Python scripts that convert CSV exports into compact indexes without losing fields.

The essence of the problem: missing recommended combinations and the total cost of the operation usually stems from the fact that this data lies in a different chunk than the main service, so a separate relationship index here will provide more accurate answers.

How does the total cost of a surgical case differ - is it a fixed amount in the database or the sum of the base price plus ancillary services? And where exactly is the context currently breaking: in the size of the Markdown files or in the logic of SKILL.md?

Yes, I am ready to sign an NDA. Let's have a call, I will analyze your task technically for free and draft a scheme for the new knowledge base architecture before the call.

Budget: 5000 UAH Deadline: 14 days

Hello. There is already a similar implemented system for your task in my cases. Take a look. It can be slightly adapted to your specifics. I will complete it in 2 weeks.

Budget: 5000 UAH Deadline: 7 days

I see that the main pain point is not in the prompt, but in fragmentation: the service is found, but the recommended combinations, prices by subdivisions, and the total cost of the surgical case get lost somewhere between the indexes. This is a classic problem of ChatGPT Projects — retrieval pulls the first relevant block and stops, not reading the connections.

I will start the diagnosis precisely from this gap. I will collect a benchmark set of queries like "cholecystectomy" — with mandatory fields: code, base price, total case cost, subdivisions, performers, packages. I will run it through the current SKILL.md and see exactly where the context breaks — in the structure of Markdown files, in the routing of sources, or in the logic of mandatory search.

⚙️ The key idea of fixation: to rebuild the generator from CSV so that one service code = one self-sufficient record.

Price, subdivisions, performers, preparation, packages, and combinations must be stitched into one block — the model should not have to gather them from different ends of a large file. Separately — a deterministic lookup for codes and prices, semantic search for conversational queries from operators. The assistant should not invent prices or transfer data between similar services — this needs to be embedded in the search rules, rather than relying on the model's "obedience."

From experience — I have built RAG assistants with structured databases of thousands of positions, written Python pipelines to convert CSV into compact indexes, and configured source routing in ChatGPT Skills. I understand the limitations of Project Sources on large arrays.

✅ Plan: audit of SKILL.md and database structure → rewritten SKILL.md with strict search steps → optimized Markdown/JSONL files → refined Python generator → set of control queries with evaluation criteria → testing on real scenarios → final documentation for support after CSV updates.

I am ready to sign an NDA. Relevant examples will be sent in the Freelancehunt chat after your response. If okay — I will send a brief plan for the first stage of the audit, and we will start. Expand

Budget: 4750 UAH Deadline: 7 days

Good day, I have over 5 years of experience in development, regularly working with RAG, LLM, and OpenAI integrations, specializing in building and optimizing corporate AI assistants. I have done comprehensive AI automation for the business platform "I Love Kebab" — there I also addressed semantic search tasks, integration with large knowledge bases, and response stability. Analyzing model behavior on large Markdown files and optimizing retrieval logic is a key part of this project. In more detail: I will start the diagnostics with a review of SKILL.md, indexing schemes, and knowledge base generator control. Could you let me know if there is a test set of real operational queries for verification? Are you ready to sign an NDA? The timeline is up to a week for auditing and basic refinement; I will provide the exact cost after a few clarifications or in private messages. I look forward to your message in private!

Budget: 13000 UAH Deadline: 10 days

I have experience in building complex search systems and processing large volumes of structured data, including optimizing retrieval logic and integrating with LLM. However, this project requires deep expertise in prompt engineering and RAG architecture, which I have less experience with than backend development. Let's discuss whether I am the right choice, or if I can assist with Python scripts for processing CSV and data transformation — please describe which specific aspect of the technical audit is a priority?

Budget: 5000 UAH Deadline: 5 days

Good day!

1. Experience with ChatGPT, RAG, and corporate knowledge bases. I built my own RAG pipeline from scratch (chunked embeddings, cosine-similarity search, pluggable vector store — sqlite-vec/Pinecone) for a desktop AI assistant, including re-indexing logic to ensure outdated fragments are not duplicated and do not remain "dead weight" in the database. This is not specifically about ChatGPT Skills, but the fundamental issues are the same: why the model skips fields in large files, how the size and structure of the source affect retrieval quality, how to build indexes that do not degrade over time.

2. Examples of similar projects. The same assistant integrated four different LLM providers (Anthropic, OpenAI-compatible, Gemini, Ollama) with native tool-calling and a retrieval layer written independently, without Langchain — meaning I understand the mechanics from the inside, not just how to configure a ready-made tool.

3. How I would approach diagnosing field skips in large files. First — an audit of current SKILL.md instructions and the sizes of Markdown files: whether they exceed the effective retrieval window, beyond which the model "does not scroll" to the required field. Next — isolated testing: the same queries against a single file and against the full corpus, to distinguish the chunking problem from the prompt/instructions problem. The key hypothesis I would check first: packages, recommended combinations, and performers are likely currently in separate files, and the model does not always "remember" to check them all — here a compact mandatory index-reference helps, which is always pulled along with the main service, rather than relying on the model's memory. Finally — a set of control queries (like the example with cholecystectomy in the specifications) with clear criteria, so that regression is caught immediately after each change, rather than during live use by operators.

4. Timelines and cost. The specified budget of 5000 UAH looks more like the cost of the initial audit (points 1-2 from your task list: audit of SKILL.md, knowledge base, and search logic, analysis of data skip reasons) — 3-4 days. The full scope (rewriting SKILL.md, restructuring Markdown/indexes, refining the Python generator with CSV, test set, testing on real scenarios, documentation) would be 15,000-25,000 UAH, 7-10 working days, depending on how deeply the generator needs to be reworked. Please clarify: is 5000 UAH the entire project or just the initial diagnostic phase?

5. NDA. Yes, I am ready to sign a non-disclosure agreement. Expand

Budget: 5000 UAH Deadline: 5 days

Yevhen, the problem here is not only similar to a prompt but also that critical fields are broken between large Markdown sources, and retrieval finds the main service but does not pull in the links, prices, or case costs.

I have practical experience with RAG and LLM response verification. In the Telegram RAG bot youtube_helper, I combined FAISS, exact keyword lookup, and MMR, and in lexai, I built grounded search for court decisions. I also write Python pipelines for structured data and systematically test models on control sets.

I will start with a benchmark set of real queries and a table of required fields. Then I will check which fragments actually fit into the context and separate the search: semantics for imprecise queries, deterministic lookup for codes, prices, subdivisions, and performers. After that, I will optimize the CSV → Markdown/JSONL generator and add regression tests.

The rate of 5000 UAH and 5 days covers the audit, diagnostics, new index scheme, first revision of SKILL.md/generator, and control runs. The full scope will be fixed after the audit. I will sign the NDA.

Please let me know the total size of Project Sources and whether logs or examples of queries are stored where fields are currently being missed.

Budget: 5000 UAH Deadline: 1 day

Hello! I have experience with LLM, RAG, and prompt engineering (ChatGPT). I would be happy to help you!

Budget: 9000 UAH Deadline: 5 days

Hello!

Judging by the description, the main issue is not with the SKILL.md itself, but rather that retrieval is losing data at the intersection of several sources — the model finds the main service but does not pull in related packages, combinations, or the total cost of the surgical case, which is a classic sign of poor context fragmentation and lack of strict routing between indexes.

I have practical experience in building RAG agents and working with structured knowledge bases based on large CSV/Markdown arrays, particularly with mandatory validation of responses against the source without data conjecture — a principle that is critical for your case with pricing and medical indications.

I will start with an audit of the current SKILL.md and the file structure in Project Sources, identifying specific patterns of field omission with real examples like "cholecystectomy," after which I will redesign the indexes so that packages, combinations, and operation costs are pulled in forcefully rather than relying on luck in searching; I will prepare a set of control queries separately to assess the quality of responses and check the Python knowledge base generator with CSV.

The audit with diagnostics of the causes and a plan for corrections will be ready in 3-4 days, with an estimated cost of 9000 UAH depending on the volume of files, and complete refinement of SKILL.md and indexes will be a separate stage after agreeing on the terms of reference; I will sign the NDA without questions.

Can I take a look at the current SKILL.md and the structure of 2-3 typical Markdown files to provide a more accurate estimate of the timelines? Expand

Budget: 10000 UAH Deadline: 7 days

I can solve your problem and simply transfer you to Claude, as chatGPT has a smaller context window.

Budget: 5000 UAH Deadline: 10 days

Hello! I have reviewed your task. The task is completely clear: it is necessary to eliminate data gaps during retrieval, optimize the structure of Markdown indexes, rewrite SKILL.md, and set up stable extraction of related services, bundles, and case costs. I have practical experience working with LLM, RAG, prompt engineering, and data processing in Python, as well as in medicine, particularly with MIS. I developed and optimized the system salesslon.com, where RAG is implemented specifically for the effective and accurate operation of LLM with large volumes of structured and unstructured data, eliminating hallucinations and ensuring precise search through deep indexes. I will approach the diagnostics and problem-solving as follows:
1. I will conduct an audit of the current SKILL.md and analyze how ChatGPT performs retrieval based on the existing Markdown structure.
2. I will focus on context fragmentation. Large Markdown files often cause loss of details (lost in the middle), so they need to be broken down into smaller, semantically dense indexes with clear meta-markup (for example, separate maps for the relationships "service - combination/package" and "operation - total cost").
3. I will optimize Python scripts that generate databases from CSV exports of MIS, so that the output data structure is maximally adapted for ChatGPT Projects search algorithms.
4. I will set up a control dataset of queries to test the stability of responses. The estimated time for the audit and initial optimization is 3-5 days. The final cost and stages will be clarified after reviewing the current database files and SKILL.md. I am fully ready to sign an NDA and adhere to confidentiality conditions. I look forward to collaborating!

Budget: 5000 UAH Deadline: 7 days

Hello!

Judging by the description, this is more of a retrieval boundary in a large database than a prompt issue. When service records, combinations, prices by departments, and performers are stored in large Markdown files, the search often finds the main service but does not pull related fields because they ended up in a different context fragment or in a source with lower priority. Therefore, "cholecystectomy" returns the basic operation without the total case cost.

My experience for this task:
- I built a RAG assistant where semantic search was combined with exact search by codes and key fields. The exact lookup prevented the model from missing critical data such as price or code. Stack: OpenAI embeddings, vector search, PostgreSQL/Supabase, separate indexes.
- I regularly convert CSV and structured exports into clean Markdown/JSON via Python, so the database generator from your CSV is clear to me.

How I diagnose field omissions: I take real operator queries and see which fragment actually pulled through and where the record broke during chunking. I check whether prices, combinations, and case costs are at all included in one selected block, and whether sources in Project Sources compete for priority. Critical fields (code, price, department, performer, total cost) I derive from deterministic search, rather than relying on the model to mention them on its own. Plus, I collect a control set of queries with expected responses so that changes can be measured rather than estimated visually.

Timeline and cost: I suggest starting with a paid audit as a separate stage (analysis of SKILL.md, database structure and search logic, report on the reasons for omissions, refinement plan, and a set of control queries). Audit: [timeline] and [amount]. The scope of the actual refinement I will specify more accurately after the audit, as it depends on what the diagnostics show. This aligns with your approach to agree on the final cost after the terms of reference.

I am ready to sign an NDA. I only work with the reference database of services and prices; I do not need personal patient data.

The assistant should respond in Ukrainian, and I will be testing it with Ukrainian operator queries, including short and conversational ones.

I am ready to show my RAG case and the first draft plan for the audit in more detail. Write to me, and we will discuss the details. Expand

Budget: 5000 UAH Deadline: 4 days

Hello, Yevhen! Very interesting and professionally formulated task. The problem of data loss during retrieval in large Markdown files and the embedded RAG in ChatGPT is a classic case of excessive or complete chunking.

I am Nina, the manager of the IT team at Valflow. The architecture of the knowledge base, optimization of SKILL.md, and Python processing scripts will be handled by Valentyn. He is a Senior Python & AI Engineer who regularly builds and configures RAG systems, semantic search, and chatbots for businesses.

Answers to your questions from the terms of reference:

1. Experience with ChatGPT, RAG, and knowledge bases:
We have extensive practical experience in developing corporate assistants, designing custom RAG systems, and optimizing context for GPT 4o. We are well-versed in the mechanics of vector search, context window limitations, and the specifics of ChatGPT Projects and Custom GPTs.

2. How we will approach diagnosing field loss:
* Markdown issue: large text files are perceived by models as a solid canvas, causing edge data and nested fields to be lost.
* Solution: we will convert the references into a compact strict JSONL or structured key-value Markdown with atomic separation.
* Two-layer search (Routing): we will create a lightweight navigator index (Master Index) that first determines the exact category and service, and then directs the assistant to a specific narrow source file.
* Optimization of the Python generator: we will rewrite the script so that each service receives unique tags, synonyms, and clear mandatory links to packages, doctors, and prices.

3. Optimization of SKILL.md:
We will rewrite the system prompt based on the principles of Strict Guardrails and Chain of Thought, requiring the model to check for the presence of the complex and total cost of the operation before providing the final answer.

4. NDA:
Yes, we are fully prepared to sign a non-disclosure agreement (NDA).

The deadline for the audit and setup: 4-6 days.
Estimated cost: 5,000 UAH.

We are ready to start auditing your SKILL.md and current scripts immediately after agreement! Expand

Budget: 4999 UAH Deadline: 6 days

Hello! I have practical experience with ChatGPT, RAG, knowledge bases, and prompt engineering. I have worked on similar projects and with structured Markdown/CSV data. To diagnose missing fields, I will check the quality of chunks, search settings, and logic in SKILL.md. Cost: 5000 UAH. Deadline: 7 days. I am ready to sign a non-disclosure agreement (NDA). I am ready to start the audit!

Budget: 5000 UAH Deadline: 7 days

Hello, Yevhen.

I see a critical risk not only in the large Markdown files but also in the fact that the data for one service is scattered among prices, packages, combinations, performers, and surgical cases. I will start with a control set of queries with reference answers and a matrix of mandatory fields — this will show exactly where the information is lost: during the generation of CSV → Markdown, fragmentation of sources, or routing of searches.

Next, I will propose a structure where each service has a self-sufficient record and a compact index of relationships; I will update SKILL.md and the Python generator and add regression checks after each CSV update.

I do not have a public medical RAG case, so I suggest starting with a separate audit with a reproducible report and demonstration on your control scenarios. I am ready to sign an NDA.

Budget: 5000 UAH Deadline: 10 days

Good day,
My experience in Python, API integrations, processing large datasets, and ETL processes aligns well with your AI assistant optimization project.

I specialize in:
• Auditing and redesigning knowledge bases to ensure reliable search in large repositories
• Developing Python scripts for transforming CSV→Markdown and indexing structured data
• Designing knowledge architecture that eliminates context fragmentation and gaps in critical fields
• Testing answer quality on real scenarios

My diagnostic approach:
1. Analyzing the current SKILL.md logic and structure of Markdown files
2. Testing real queries (e.g., "cholecystectomy") to identify retrieval failure points
3. Mapping coverage: services → packages → combinations → prices → performers
4. Optimizing the Python generator and restructuring indexes for guaranteed search of all required fields

Estimated: 1-2 weeks 4500–5500 UAH (depending on the database volume).
I am ready to sign an NDA and provide detailed documentation for further support.

I look forward to the Technical Assignment. Expand

Budget: 5000 UAH Deadline: 5 days

Good day! We have experience in setting up RAG systems and optimizing prompts for medical services. We implement this through an audit of the current architecture, improving contextual search, and introducing new skills to enhance the accuracy of the assistant's responses. We are ready to begin analyzing your current knowledge base and settings.

Budget: 18000 UAH Deadline: 5 days

Hello!

I have over 10 years of experience in Full-Stack development and practical experience in building LLM functionality, structured knowledge bases, multi-step AI scenarios, and systems where the model must work only with verified data.

In my own products, I have worked with:

— LLM API and structured output;
— large subject models and catalogs;
— searching for related entities;
— data normalization;
— context generation and verification;
— Python scripts for data preparation;
— rules that prevent the model from inventing missing information.

In your case, I would start not with rewriting prompts, but with diagnosing the entire retrieval process.

The main reasons for gaps that I would check first:

— overly large Markdown files and unsuccessful fragmentation boundaries;
— splitting related fields between different fragments;
— ambiguous service names and lack of normalized synonyms;
— competition between main services, packages, and recommendations during search;
— lack of a separate search route for operations;
— excessive dependence on a single semantic query;
— indexes that point to a file but do not guarantee retrieval of all mandatory fields.

For a query like "cholecystectomy," I would implement a separate scenario:

1. Normalization of the query and search for synonyms.
2. Search for relevant service codes.
3. Mandatory retrieval of prices for each subdivision.
4. Separate search for the cost of the surgical case.
5. Search for performers and equipment.
6. Search for packages, complexes, and recommended combinations.
7. Verification of the completeness of the response before showing it to the operator.
8. Prohibition of outputting fields that are not verified by the source.

Probably, the optimal solution would not be one large Markdown database, but a combination of:

— compact indexes for routing;
— separate files by entity types;
— stable identifiers for data linking;
— JSONL or compact structures for critical fields;
— Markdown for readable reference descriptions;
— a control layer that checks for the presence of all mandatory fields.

I will also prepare a set of control queries: exact, shortened, conversational, with errors, with synonyms, and ambiguous names. For each scenario, specific criteria can be defined: correctness of code, prices, subdivision, performer, package, total cost of the operation, and absence of unverified data.

I propose to divide the work into two stages:

**Stage 1 — audit and prototype of corrections**
Analysis of SKILL.md, source structure, generator, control scenarios, and reasons for gaps.

Approximately: **3–5 working days** Expand

Budget: 5000 UAH Deadline: 1 day

Hello. I have significant experience in developing and optimizing RAG systems for large corporate knowledge bases, particularly in the medical field. My approach will include a thorough audit of the current knowledge base architecture, indexing, and chunking strategies to systematically eliminate issues with data loss and response instability. I will focus on optimizing SKILL.md, enhancing prompt engineering logic, and Python scripts for source formation, ensuring accurate extraction of all necessary fields and consistency of responses. I have developed templates for vectorization and multi-stage RAG processes that will expedite the implementation of solutions and allow for quick stabilization of the system. I propose to discuss all implementation details, final budget, and timelines in private messages.

Budget: 5000 UAH Deadline: 7 days

A stable assistant that does not miss recommended combinations and always extracts the full cost of a surgical case — this is achieved through the correct architecture of indexes and prompt routing in ChatGPT Projects.

I built RAG systems on OpenAI for corporate catalogs: I encountered context fragmentation when a large Markdown file is split in such a way that individual fields end up outside the retrieval window. The first thing I will do is audit the size of chunks, the anchor structure of files, and the order of connecting sources in SKILL.md.

Next: I will restructure files according to Project Sources limits, set up mandatory multi-source searches for packages, combinations, performers, and surgeries, update the Python CSV generator to MD, create a test set based on real scenarios, and document the logic for future updates.

I am ready to sign an NDA. The estimated volume is 25-30 hours: audit, optimization, testing, documentation.

Write to me — we will discuss the details and clarify the terms of reference.

Budget: 5000 UAH Deadline: 4 days

Good day, I am a Python developer. I have experience in developing AI agents using RAG databases, graphs, LangGraph, LangChain, etc., for tracking response accuracy, fine-tuning, and providing answers. Write to me - we will discuss the details. I am ready to start today.

Budget: 5000 UAH Deadline: 1 day

Hello. Evgen Galinskiy.

Thanks to my extensive experience, high level of responsiveness, and code quality, I will perfectly implement the current requirements.

After further discussion, I can provide a specific deadline and budget.

Budget: 5000 UAH Deadline: 4 days

I looked at the technical specifications: large MD files overwhelm the context, the model loses packages, combinations, operation costs. I will break down the indexes separately: services, packages, combinations, prices, performers. I will rewrite SKILL.md with mandatory search, and check the Python CSV generator. Test on "cholecystectomy." Ready in 4 days.

Budget: 5000 UAH Deadline: 5 days

The omission of the cost of the surgical case and recommended combinations, as in the example with cholecystectomy, occurs due to excessively large Markdown files in Project Sources, the model does not reach the required field. I would break the files into compact indexes by types: services, packages, combinations, surgery, and specify mandatory routing in SKILL.md with control queries for verification. For auditing and refinement, approximately 5 days, I will sign the NDA.