Vectara icon
developer-tools

Vectara

Vectara gives you a grounded-answer API with strong refusals, but tables, contradictions, scans, and pricing are the weak spots.

Visit Vectara
10/10 refusals4/8 tablesScanned PDFs failSales-gated pricing
TL;DR — our verdictUpdated August 2026 · 15 test artifacts

Strong grounding, but the hard cases are uneven.

Where it wins
  • You need a grounded-answer API that refuses unsupported questions instead of inventing answers.
  • You can tolerate weaker table-cell and contradiction performance if refusal discipline matters more.
  • You need native pipeline observability and verified tenant isolation.
Main limitation
  • You need scanned-PDF OCR.
Pricing (verified plans)
30 Day Free Trial FreeSaaS Starting at $100K/yearVPC Starting at $250K/yearOn-prem Starting at $500K/year
Strongest test artifacts
Corpus-Grounded Question Answering →Document Ingestion →Web Crawl Pipelines →

Our take

Vectara's refusal discipline and citation fidelity were excellent, and freshness, isolation, and pipelines all looked real. But table cells, contradictory editions, scanned PDFs, and native CSV ingest were all weak spots, and paid pricing remains sales-gated.

Download video: Demo videoWatch video (streaming): Demo video
Narrated session recording covering login, corpora and documents views, live io queries, and the scanned-PDF upload failure.

In-Depth Review

Our detailed analysis of Vectara — features, performance, and real-world testing.

AD
AI Demos Team
Expert Reviewer
Verified Review

Feature-by-Feature Breakdown

Corpus-Grounded Question Answering
Usable, but not consistent.
10/10
▾
Test Summary
Feature tested: Corpus-Grounded Question Answering
Result: Failed (10/10) — Usable, but not consistent.

Feature tested: Corpus-Grounded Question Answering

Result: Failed (10/10)

Verdict: Usable, but not consistent.

Expected behavior: Vectara answers in-corpus factual questions with cited support and can refuse unsupported questions. The tested prompts ranged from direct factual queries to PDF table lookups and questions over conflicting document versions.

Test case: Text prompt → Image

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Image): The io workspace answered the license question directly and showed live execution steps for the corpus-backed query. — final_04_io_direct_factual.png

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Image): The io workspace answered the license question directly and showed live execution steps for the corpus-backed query. — final_04_io_direct_factual.png

What changed: Text prompt transformed into Image

Test case: Text prompt → Image

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Image): The assistant refused to invent a FY2027-28 projection because the corpus only provided the current RBI policy repo rate. — final_04_io_refusal.png

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Image): The assistant refused to invent a FY2027-28 projection because the corpus only provided the current RBI policy repo rate. — final_04_io_refusal.png

What changed: Text prompt transformed into Image

Test case: Text prompt → Image

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Image): The io workspace returned a yen-denominated figure for a table-cell question in the financial report. — final_04_io_table_numeric.png

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Image): The io workspace returned a yen-denominated figure for a table-cell question in the financial report. — final_04_io_table_numeric.png

What changed: Text prompt transformed into Image

Test case: Text prompt → Image

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Image): The response drifted to a 2026 value even though the prompt asked for the current limit, matching the contradiction failure noted in the report. — final_04_io_contradiction.png

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Image): The response drifted to a 2026 value even though the prompt asked for the current limit, matching the contradiction failure noted in the report. — final_04_io_contradiction.png

What changed: Text prompt transformed into Image

Test case: Text prompt → Text prompt

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Text prompt): Output

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Text prompt): Output

What changed: Text prompt transformed into Text prompt

Why it matters / Conclusion: Direct factual answering was often correct and the sampled citations supported the claims, but the same stack still produced false refusals on some answerable inputs.

Vectara answers in-corpus factual questions with cited support and can refuse unsupported questions. The tested prompts ranged from direct factual queries to PDF table lookups and questions over conflicting document versions.

INPUT
Using the rag-eval-corpus-api corpus, answer: Under what license is Qwen 3.8 27B released?
↓
OUTPUT
Output artifact for "Corpus-Grounded Question Answering" test: The io workspace answered the license question directly and showed live execution steps for the corpus-backed query., final_04_io_direct_factual.png
The io workspace answered the license question directly and showed live execution steps for the corpus-backed query.
INPUT
Using the rag-eval-corpus-api corpus, answer: What was the RBI repo rate projection for FY2027-28?
↓
OUTPUT
Output artifact for "Corpus-Grounded Question Answering" test: The assistant refused to invent a FY2027-28 projection because the corpus only provided the current RBI policy repo rate., final_04_io_refusal.png
The assistant refused to invent a FY2027-28 projection because the corpus only provided the current RBI policy repo rate.
INPUT
Using the rag-eval-corpus-api corpus, answer: What were total assets as of March 31, 2025?
↓
OUTPUT
Output artifact for "Corpus-Grounded Question Answering" test: The io workspace returned a yen-denominated figure for a table-cell question in the financial report., final_04_io_table_numeric.png
The io workspace returned a yen-denominated figure for a table-cell question in the financial report.
INPUT
Using the rag-eval-corpus-api corpus, answer: What is the current Social Security wage base limit for employer withholding?
↓
OUTPUT
Output artifact for "Corpus-Grounded Question Answering" test: The response drifted to a 2026 value even though the prompt asked for the current limit, matching the contradiction failure noted in the report., final_04_io_contradiction.png
The response drifted to a 2026 value even though the prompt asked for the current limit, matching the contradiction failure noted in the report.
INPUT
INPUT: Deep-in-long-doc query from the 100+ page Fed report corpus.
↓
OUTPUT
The report says deep-in-long-doc questions scored 3/5 overall, so long-document retrieval worked but was not consistently reliable.
Bottom Line
Direct factual answering was often correct and the sampled citations supported the claims, but the same stack still produced false refusals on some answerable inputs.
From our researchUpload Your Docs, Get a Grounded Answer API — RAG-as-a-Service Platformsearlier research
Document Ingestion
Good for text-bearing files; scans and CSV are the problem.
0/5
▾
Test Summary
Feature tested: Document Ingestion
Result: Failed (0/5) — Good for text-bearing files; scans and CSV are the problem.

Feature tested: Document Ingestion

Result: Failed (0/5)

Verdict: Good for text-bearing files; scans and CSV are the problem.

Expected behavior: Vectara ingests text-bearing files into corpora, including PDF, markdown, HTML, DOCX, and a Hindi PDF. The same ingest path also exposed failures for scanned/image-only PDFs and unreliable CSV handling.

Test case: Text prompt → Image

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Image): The documents list shows the uploaded corpus files present in Vectara, including the mixed-format documents. — 12_documents_list.png

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Image): The documents list shows the uploaded corpus files present in Vectara, including the mixed-format documents. — 12_documents_list.png

What changed: Text prompt transformed into Image

Test case: PDF document → Image

Input type: PDF document

Input used: Input artifact (PDF document): Input — 02_scanned_mountain-pine-beetle_USDA-1983.pdf

Observed output: Output artifact (Image): The scanned PDF upload failed with HTTP 415, and no OCR fallback was available. — vectara_console_scanned_pdf_ingestion_failed.png

Input artifact: Input artifact (PDF document): Input — 02_scanned_mountain-pine-beetle_USDA-1983.pdf

Output artifact: Output artifact (Image): The scanned PDF upload failed with HTTP 415, and no OCR fallback was available. — vectara_console_scanned_pdf_ingestion_failed.png

What changed: PDF document transformed into Image

Test case: Text/code file → Text/code file

Input type: Text/code file

Input used: Input artifact (Text/code file): Input — 04c_mixed_sheet_usgs-earthquakes.csv

Observed output: Output artifact (Text/code file): Native CSV upload was rejected with a 415-style mime-type error in the ingestion saga. — 04_ingestion_csv_saga.txt

Input artifact: Input artifact (Text/code file): Input — 04c_mixed_sheet_usgs-earthquakes.csv

Output artifact: Output artifact (Text/code file): Native CSV upload was rejected with a 415-style mime-type error in the ingestion saga. — 04_ingestion_csv_saga.txt

What changed: Text/code file transformed into Text/code file

Test case: Text/code file → Image

Input type: Text/code file

Input used: Input artifact (Text/code file): Input — 04c_mixed_sheet_usgs-earthquakes.csv

Observed output: Output artifact (Image): The workaround path succeeded once and showed the CSV content indexed as a text file. — vectara_console_csv_as_txt_upload_succeeded.png

Input artifact: Input artifact (Text/code file): Input — 04c_mixed_sheet_usgs-earthquakes.csv

Output artifact: Output artifact (Image): The workaround path succeeded once and showed the CSV content indexed as a text file. — vectara_console_csv_as_txt_upload_succeeded.png

What changed: Text/code file transformed into Image

Test case: PDF document → Text/code file

Input type: PDF document

Input used: Input artifact (PDF document): Input — 02_scanned_mountain-pine-beetle_USDA-1983.pdf

Observed output: Output artifact (Text/code file): The scanned, image-only PDF hard-failed ingestion with HTTP 415 and no OCR fallback. — vectara_api_scanned_pdf_ingestion_415_failure.txt

Input artifact: Input artifact (PDF document): Input — 02_scanned_mountain-pine-beetle_USDA-1983.pdf

Output artifact: Output artifact (Text/code file): The scanned, image-only PDF hard-failed ingestion with HTTP 415 and no OCR fallback. — vectara_api_scanned_pdf_ingestion_415_failure.txt

What changed: PDF document transformed into Text/code file

Test case: Text/code file → Image

Input type: Text/code file

Input used: Input artifact (Text/code file): INPUT: Same CSV bytes renamed to .txt and uploaded with explicit type=text/plain. — 04c_mixed_sheet_usgs-earthquakes.csv

Observed output: Output artifact (Image): The disguised .txt workaround succeeded in this attempt, but the report says the same trick only worked 2 out of 6 total attempts. — retry_csv_as_txt_v2.png

Input artifact: Input artifact (Text/code file): INPUT: Same CSV bytes renamed to .txt and uploaded with explicit type=text/plain. — 04c_mixed_sheet_usgs-earthquakes.csv

Output artifact: Output artifact (Image): The disguised .txt workaround succeeded in this attempt, but the report says the same trick only worked 2 out of 6 total attempts. — retry_csv_as_txt_v2.png

What changed: Text/code file transformed into Image

Test case: Text/code file → Text/code file

Input type: Text/code file

Input used: Input artifact (Text/code file): INPUT: 3x confirmation retest of the disguised CSV upload. — 04c_mixed_sheet_usgs-earthquakes.csv

Observed output: Output artifact (Text/code file): The retest log confirms the non-deterministic behavior and the unexpected deduplication behavior. — 06_csv_nondeterminism_3x.txt

Input artifact: Input artifact (Text/code file): INPUT: 3x confirmation retest of the disguised CSV upload. — 04c_mixed_sheet_usgs-earthquakes.csv

Output artifact: Output artifact (Text/code file): The retest log confirms the non-deterministic behavior and the unexpected deduplication behavior. — 06_csv_nondeterminism_3x.txt

What changed: Text/code file transformed into Text/code file

Why it matters / Conclusion: Clean text-bearing files loaded well, but scanned PDFs failed outright and CSV support was not dependable; the Hindi PDF ingested cleanly, though cross-lingual retrieval later struggled on some labeled questions.

Vectara ingests text-bearing files into corpora, including PDF, markdown, HTML, DOCX, and a Hindi PDF. The same ingest path also exposed failures for scanned/image-only PDFs and unreliable CSV handling.

INPUT
Upload the standard mixed-format corpus (PDF, markdown, HTML, DOCX, and Hindi PDF) into the same rag-eval corpus.
↓
OUTPUT
Output artifact for "Document Ingestion" test: The documents list shows the uploaded corpus files present in Vectara, including the mixed-format documents., 12_documents_list.png
The documents list shows the uploaded corpus files present in Vectara, including the mixed-format documents.
INPUT
02_scanned_mountain-pine-beetle_USDA-1983.pdf
↓
OUTPUT
Output artifact for "Document Ingestion" test: The scanned PDF upload failed with HTTP 415, and no OCR fallback was available., vectara_console_scanned_pdf_ingestion_failed.png
The scanned PDF upload failed with HTTP 415, and no OCR fallback was available.
INPUT
04c_mixed_sheet_usgs-earthquakes.csv
Loading file...
↓
OUTPUT
04_ingestion_csv_saga.txt
Loading file...
Native CSV upload was rejected with a 415-style mime-type error in the ingestion saga.
INPUT
04c_mixed_sheet_usgs-earthquakes.csv
Loading file...
↓
OUTPUT
Output artifact for "Document Ingestion" test: The workaround path succeeded once and showed the CSV content indexed as a text file., vectara_console_csv_as_txt_upload_succeeded.png
The workaround path succeeded once and showed the CSV content indexed as a text file.
INPUT
02_scanned_mountain-pine-beetle_USDA-1983.pdf
↓
OUTPUT
vectara_api_scanned_pdf_ingestion_415_failure.txt
Loading file...
The scanned, image-only PDF hard-failed ingestion with HTTP 415 and no OCR fallback.
file
04c_mixed_sheet_usgs-earthquakes.csv
Loading file...
INPUT: Same CSV bytes renamed to .txt and uploaded with explicit type=text/plain.
↓
OUTPUT
Output artifact for "Document Ingestion" test: The disguised .txt workaround succeeded in this attempt, but the report says the same trick only worked 2 out of 6 total attempts., retry_csv_as_txt_v2.png
The disguised .txt workaround succeeded in this attempt, but the report says the same trick only worked 2 out of 6 total attempts.
file
04c_mixed_sheet_usgs-earthquakes.csv
Loading file...
INPUT: 3x confirmation retest of the disguised CSV upload.
↓
OUTPUT
06_csv_nondeterminism_3x.txt
Loading file...
The retest log confirms the non-deterministic behavior and the unexpected deduplication behavior.
Bottom Line
Clean text-bearing files loaded well, but scanned PDFs failed outright and CSV support was not dependable; the Hindi PDF ingested cleanly, though cross-lingual retrieval later struggled on some labeled questions.
From our researchUpload Your Docs, Get a Grounded Answer API — RAG-as-a-Service Platformsearlier research
Multi-Tenant Corpus Isolation
Solid.
▾
Test Summary
Feature tested: Multi-Tenant Corpus Isolation
Result: Passed — Solid.

Feature tested: Multi-Tenant Corpus Isolation

Result: Passed

Verdict: Solid.

Expected behavior: Vectara keeps separate corpora isolated so questions answerable only from another tenant's documents are refused instead of leaking across tenants. The hand test reported no cross-tenant leakage.

Test case: Text prompt → Text prompt

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Text prompt): Output

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Text prompt): Output

What changed: Text prompt transformed into Text prompt

Test case: Text prompt → Text prompt

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Text prompt): Output

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Text prompt): Output

What changed: Text prompt transformed into Text prompt

Why it matters / Conclusion: Isolation looked strong in the hand test: the second tenant stayed separate and the answer was correctly refused rather than leaked from the other corpus.

Vectara keeps separate corpora isolated so questions answerable only from another tenant's documents are refused instead of leaking across tenants. The hand test reported no cross-tenant leakage.

INPUT
INPUT: Ask a question answerable only from documents that were not loaded into rag-eval-tenant2.
↓
OUTPUT
Correctly refused with zero cross-tenant leakage; the second corpus contained only 2 of the 11 source documents.
INPUT
Second tenant corpus containing only documents 1 and 3; ask a question answerable only from a document absent from that tenant.
↓
OUTPUT
The query was correctly refused and no document from the other tenant leaked into the result set.
Bottom Line
Isolation looked strong in the hand test: the second tenant stayed separate and the answer was correctly refused rather than leaked from the other corpus.
From our researchearlier researchUpload Your Docs, Get a Grounded Answer API — RAG-as-a-Service Platforms
Web Crawl Pipelines
Functional but not turnkey.
▾
Test Summary
Feature tested: Web Crawl Pipelines
Result: Partial — Functional but not turnkey.

Feature tested: Web Crawl Pipelines

Result: Partial

Verdict: Functional but not turnkey.

Expected behavior: Vectara provides a native pipelines workflow for web-crawl jobs with live run status and observability. The evaluated flow required an explicit agent choice, and the crawled output destination was unclear.

Test case: Text prompt → Image

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Image): The Pipelines page started empty and offered a Create pipeline button. — connector_01_pipelines_list.png

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Image): The Pipelines page started empty and offered a Create pipeline button. — connector_01_pipelines_list.png

What changed: Text prompt transformed into Image

Test case: Text prompt → Image

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Image): The wizard required an agent selection and exposed multiple agent choices. — connector_11_agent_dropdown.png

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Image): The wizard required an agent selection and exposed multiple agent choices. — connector_11_agent_dropdown.png

What changed: Text prompt transformed into Image

Test case: Text prompt → Image

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Image): The run showed live progress, a 100% success rate, total records, processed records, and zero failed records. — connector_18_run_progress.png

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Image): The run showed live progress, a 100% success rate, total records, processed records, and zero failed records. — connector_18_run_progress.png

What changed: Text prompt transformed into Image

Test case: Text prompt → Text/code file

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Text/code file): A fresh API check found no new corpus and no new documents, so the crawl destination remained unclear. — vectara_api_postcrawl_corpus_verification.txt

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Text/code file): A fresh API check found no new corpus and no new documents, so the crawl destination remained unclear. — vectara_api_postcrawl_corpus_verification.txt

What changed: Text prompt transformed into Text/code file

Why it matters / Conclusion: The pipeline stack is real and observable, but the required agent step and unclear output destination make the flow feel incomplete.

Vectara provides a native pipelines workflow for web-crawl jobs with live run status and observability. The evaluated flow required an explicit agent choice, and the crawled output destination was unclear.

INPUT
Open the Pipelines area before creating a new web-crawl workflow.
↓
OUTPUT
Output artifact for "Web Crawl Pipelines" test: The Pipelines page started empty and offered a Create pipeline button., connector_01_pipelines_list.png
The Pipelines page started empty and offered a Create pipeline button.
INPUT
Choose an agent on the Transform step of the pipeline wizard.
↓
OUTPUT
Output artifact for "Web Crawl Pipelines" test: The wizard required an agent selection and exposed multiple agent choices., connector_11_agent_dropdown.png
The wizard required an agent selection and exposed multiple agent choices.
INPUT
Run the crawl pipeline and watch the run history.
↓
OUTPUT
Output artifact for "Web Crawl Pipelines" test: The run showed live progress, a 100% success rate, total records, processed records, and zero failed records., connector_18_run_progress.png
The run showed live progress, a 100% success rate, total records, processed records, and zero failed records.
INPUT
Verify where the crawled content landed after the run.
↓
OUTPUT
vectara_api_postcrawl_corpus_verification.txt
Loading file...
A fresh API check found no new corpus and no new documents, so the crawl destination remained unclear.
Bottom Line
The pipeline stack is real and observable, but the required agent step and unclear output destination make the flow feel incomplete.
From our researchearlier researchUpload Your Docs, Get a Grounded Answer API — RAG-as-a-Service Platforms

Public pricing

Only the trial is self-serve; the paid tiers are sales-gated.

30 Day Free Trial
Free
Self-serve trial
SaaS
Starting at $100K/year
Contact sales
VPC
Starting at $250K/year
Contact sales
On-prem
Starting at $500K/year
Contact sales

Captured from Vectara's public pricing page; no usage-based public rate card was shown.

✓ Use This If
●You need a grounded-answer API that refuses unsupported questions instead of inventing answers.
●You can tolerate weaker table-cell and contradiction performance if refusal discipline matters more.
●You need native pipeline observability and verified tenant isolation.
✕ Skip This If
●You need scanned-PDF OCR.
●You need dependable native CSV ingestion.
●You need contradiction-aware selection of current editions.
●You need a public self-serve usage-price card.
●You need crawl output to land transparently in a corpus without extra wiring.
developer-toolsapistextFounderOther
Yes. In the refusal test it correctly refused all 10/10 out-of-corpus questions across three rounds, with zero cited hallucinations.
No. The scanned/image-only PDF hard-failed ingestion with HTTP 415 and the report found no OCR fallback in the API or console.
Mixed. Table/numeric retrieval scored 4/8, so it can answer some table-cell questions but misses others even when the number is present in the document.
In the scored contradiction set, Vectara resolved every question to the superseded edition instead of surfacing the conflict or preferring the newer file.
Not reliably. Native .csv uploads were rejected, and a renamed .txt workaround only succeeded inconsistently across repeated attempts.
The public pricing page showed a self-serve 30-day free trial and sales-gated paid tiers: Scale starting at $100K/year, VPC at $250K/year, and On-prem at $500K/year.
The crawl ran and was observable, but a fresh post-crawl API check found no new corpus or document growth, so the destination wiring was still unclear.
The reported latency on a 144-query run was p50 1.89s, p95 2.67s, and p99 3.88s, with 0 failures after retry.

Banner Preview

How the embed badge will look on your site

Vectara featured on AI Demos

Embed HTML

Copy this code to your website source

<a target="_blank" href="https://aidemos.com/tools/vectara?utm_source=vectara_embed" style="width: 250px; height: 80px; border-radius:4px;" width="250" height="80"> <img src="https://aidemos-website-images.s3.amazonaws.com/featured.png" alt="Vectara | Featured on AI Demos" style="width: 250px; height: 80px; border-radius:4px;" width="250" height="80"> </a>

Quick Integration Guide

  • 1Copy the HTML code block above.
  • 2Paste it into your site's HTML or CMS editor.
  • 3Banner appears instantly on your page.
  • 4Links back to your tool profile here.
Similar Tools

Similar Tools

Discover more AI tools like Vectara to enhance your workflow.

🤖
Dify
API-first managed RAG with strong table answers and connectors, but you still need guardrails for scans, multilingual retrieval, and refusal cases.
AI Tool
🤖
Vectorize.io
Managed RAG API with clean ingestion, OCR, and public pricing, but with caveats on refusals and freshness.
AI Tool
🤖
RAGFlow
AI Tool
🤖
CustomGPT
A knowledge-base website chatbot that feels unusually human and keeps follow-up context well, but it can still hallucinate support contacts.
AI Tool

Comments (0)

Please Log in to join the discussion.

Built by FutureSmart AI — the team behind AI Demos

Need a custom AI solution for this use case?

If you are looking to build a custom grounded-answer assistant, RAG system, or enterprise search tool for your business or internal workflow, email us at contact@futuresmart.ai.

Get a custom build

Found something inaccurate or missing? We try to keep our AI research accurate and useful. If you found outdated information, an issue, or have a suggestion, email us at collaborate@aidemos.com.

Back to Top